阅读说明：本技术 在表征生理系统时配置和使用神经网络的方法和系统 (Method and system for configuring and using neural networks in characterizing physiological systems ) 是由 A·库索斯 T·W·F·伯顿 H·吉林斯 S·拉姆昌达尼 W·桑德斯 I·沙德弗斯 于 2019-12-23 设计创作，主要内容包括：例示方法和系统有助于通过生物物理信号数据集配置和训练神经网络或其全部(例如深度神经网络、卷积神经网络(CNN)等)或其全部,用于对受试者中疾病或病理的存在或不存在进行确定地估计并对疾病或病理进行评估和/或分类,包括例如对受试者中此类疾病或病理的严重性进行评估和/或分类。在心脏的环境下,本文描述的方法和系统有助于通过心脏信号数据集配置和训练神经网络或其全部,以对冠状动脉疾病的存在或不存在或冠状动脉病理进行确定的估计。(The example methods and systems facilitate configuration and training of a neural network, or all thereof (e.g., a deep neural network, a Convolutional Neural Network (CNN), etc.), or all thereof, through a biophysical signal dataset for deterministically estimating and assessing and/or classifying the presence or absence of a disease or pathology in a subject, including, for example, assessing and/or classifying the severity of such disease or pathology in a subject. In the context of the heart, the methods and systems described herein facilitate the configuration and training of neural networks, or all thereof, from cardiac signal datasets to make a deterministic estimate of the presence or absence of coronary artery disease or coronary artery pathology.)

1. A method, comprising:

receiving, by a processor, a set of biophysical signal data of a subject acquired from one or more channels of one or more sensors;

pre-processing the biophysical signal data set to generate one or more pre-processed data sets, wherein each pre-processed data set comprises a single isolated complete cardiac cycle; and

determining, by the processor, a value indicative of the presence of a heart disease or condition by inputting the preprocessed data set directly to one or more deep neural networks trained using a set of training biophysical signal data sets acquired from patients: the patient is diagnosed with a heart disease or condition and is marked for the presence or absence of a heart disease or condition,

wherein an output data set is output via a report and/or display based on the determined value indicative of the presence of a heart disease or condition.

2. The method of claim 1, wherein the heart disease or condition is coronary artery disease, and wherein the step of determining a value indicative of the presence of a heart disease or condition comprises:

inputting the preprocessed data sets into a set of one or more deep neural networks trained using one or more biophysical signal data sets acquired from a plurality of patients labeled with a diagnosis of the presence of coronary artery disease,

wherein an output of the one or more deep neural networks is output as the output data set via the report and/or the display.

3. The method of claim 1 or 2, wherein the biophysical signal dataset is acquired from two or more acquisition channels, and wherein the preprocessed datasets from each of the acquisition channels are phase synchronized.

4. The method of any one of claims 1-3, wherein the step of preprocessing the biophysical signal dataset comprises:

segmenting, by the processor, a portion of a biophysical signal dataset associated with a first acquisition channel of the one or more acquisition channels or a normalized dataset derived from the portion of the biophysical signal dataset into one or more first segmented datasets, wherein each of the first segmented datasets comprises a single isolated complete cardiac cycle as a first single isolated complete cardiac cycle, wherein the first single isolated complete cardiac cycle has an associated time window; and

segmenting, by the processor, another portion of the biophysical signal data set associated with a second acquisition channel of the one or more acquisition channels or a normalized data set derived from another portion of the biophysical signal data set into one or more second segmented data sets, wherein each of the one or more second segmented data sets comprises a second single isolated complete cardiac cycle, wherein the second single isolated complete cardiac cycle has an associated time window corresponding to a time window of the first single isolated complete cardiac cycle to provide a phase-synchronized data set.

5. The method of any of claims 1-4, wherein the label for marking the presence of coronary artery disease comprises a Gensini-based score determined as a combination of a severity-weighted score and a location-weighted score for coronary lesions diagnosed in the myocardium.

6. The method of claim 5, wherein the Gensini-based score is linearized.

7. The method of any of claims 1-6, further comprising:

determining, by the processor, one or more location values indicative of the presence of a heart disease or condition at a given coronary artery by inputting the preprocessed data set or a modified version of the preprocessed data set to one or more second deep neural networks trained using one or more sets of biophysical signal data acquired from a plurality of patients flagged as having a diagnosis of coronary artery disease at a coronary artery selected from the group consisting of: left aorta (LMA), proximal left circumflex (Prox LCX), Mid left circumflex (Mid LCX), distal left circumflex (Dist LCX), LPAV, first blunt edge (OM1), second blunt edge (OM2), third blunt edge (OM3), proximal left anterior descending (Prox LAD), Mid left anterior descending (Mid LAD), distal left anterior descending (Dist LAD), LAD 1, LAD 2, proximal right coronary artery (Prox RCA), Mid right coronary artery (Mid RCA), distal right coronary artery (Dist RCA) and acute marginal branch of right posterior descending (AcM RPDA),

wherein the determined one or more position values are output as the output data set via the report and/or the display.

8. The method of claim 7, further comprising:

comparing, by the processor, the value indicative of the presence of the cardiac disease or condition to a threshold,

wherein the step of determining the one or more location values indicative of the presence of a heart disease or condition at a given coronary artery is performed based on the comparison.

9. The method of any of claims 1-8, further comprising:

performing, by the processor, a phase space analysis operation on the received biophysical signal dataset or the preprocessed dataset to generate one or more phase space analysis datasets/images; and

outputting, by the processor, the one or more generated sets/images of phase space analysis data, wherein the one or more generated sets/images of phase space analysis data are presented in the report and/or the display simultaneously and/or synchronously with the output data set.

10. The method of any one of claims 1-9, wherein the step of preprocessing the biophysical signal dataset to generate one or more preprocessed datasets further comprises a second preprocessing operation selected from the group consisting of:

performing a downsampling operation;

performing a baseline wander removal operation; and

a normalization operation is performed.

11. The method of any of claims 1-10, wherein at least one of the one or more deep neural networks is configured based on a hyper-parametric search loop, wherein the hyper-parametric search loop comprises:

generating, by the processor, a plurality of hyper-parameter sets for a template convolutional neural network, wherein each of the plurality of hyper-parameter sets is generated by random or pseudo-random selection from a set of candidate hyper-parameters, at least one hyper-parameter of the set of candidate hyper-parameters being selected from the group consisting of: batch size, learning rate, convolutional layer, filter size, number of filters in the first convolutional layer, increment of filters in subsequent layers, number of additional dense layers, size of additional dense layers, activation function type, target, expansion rate, and discard;

training, by the processor, the template convolutional neural network for each of a plurality of sets of hyper-parameters, wherein in each evaluation instance, the template convolutional neural network is configured with a set of hyper-parameters of the plurality of sets of hyper-parameters; and

for each of a plurality of sets of hyper-parameters, assessing, by the processor, the trained deep neural network using a first validation dataset, wherein each assessment generates a score.

12. The method of claim 11, wherein the at least one of the one or more deep neural networks is configured based on bayesian hyper-parameter optimization.

13. The method of claim 11 or 12, wherein the assessment of the trained deep neural network comprises generating an accuracy score, a weighted accuracy score, a positive prediction score, a negative prediction score, an F score, a sensitivity score, a specificity score, and/or a diagnostic odds ratio score.

14. The method of any one of claims 7-13, wherein at least one of the one or more second deep neural networks is configured based on a hyper-parametric search loop.

15. The method of claim 2, wherein the one or more sets of biophysical signal data acquired from a plurality of patients marked with a diagnosis of the presence or absence of coronary artery disease at a coronary artery are configured as a coronary artery disease location array, and wherein the location array comprises a plurality of elements, each element corresponding to a marker indicating the presence or absence of a heart disease or condition at a given location in a coronary artery.

16. The method of any one of claims 1-15, further comprising:

modifying the value indicative of the presence of the cardiac disease or condition based on one or more additional predictive models, wherein the one or more additional predictive models include an analysis based on geometric features related to a geometry or topology of the biophysical signal dataset in phase space.

17. The method of any one of claims 1-15, further comprising:

merging the value indicative of the presence of a cardiac disease or condition with a second predictive value indicative of the presence of a cardiac disease or condition, wherein the second predictive value indicative of the presence of a cardiac disease or condition is based on one or more additional predictive models, wherein the one or more additional predictive models comprise an analysis based on geometric features related to the geometry or topology of the biophysical signal dataset in phase space.

18. A method as claimed in claim 16 or 17, wherein the geometric features relating to the geometry or topology of the biophysical signal data set in phase space comprise quantification of the biophysical signal data set in a region of phase space occupied by an identified ventricular depolarization trajectory.

19. The method of claim 18, wherein VDFA features are quantifications of fiducial points of a biophysical signal data set in phase space, wherein the fiducial points include at least one of: a machine-identified maximum ventricular depolarization, a machine-identified point prior to the maximum ventricular depolarization, and a machine-identified conclusion of ventricular depolarizations.

20. A method, comprising:

receiving, by a processor, a biophysical signal dataset of a subject, wherein the biophysical signal dataset relates to a plurality of wideband phase gradient signals acquired simultaneously from the subject via a corresponding number of acquisition channels by at least one electrode;

pre-processing a biophysical signal data set from at least one of the acquisition channels to generate one or more pre-processed data sets, wherein each pre-processed data set comprises a single isolated complete cardiac cycle; and

determining, by the processor, a value indicative of the presence or absence of a cardiac disease or condition by inputting the preprocessed data set directly into a set of one or more deep neural networks trained using one or more biophysical signal data sets acquired from a plurality of patients labeled with a diagnosis of the presence or absence of coronary artery disease;

wherein the output data set is output via the report and/or the display based on the determined values indicative of the binary presence of the cardiac disease or condition.

21. A method, comprising:

receiving, by a processor, a biophysical signal dataset of a subject, wherein the biophysical signal dataset relates to a plurality of wideband phase gradient signals acquired simultaneously from the subject via a corresponding number of acquisition channels by at least one electrode; and

determining, by the processor, one or more location values indicative of the presence or absence of a heart disease or condition at one or more coronary arteries by inputting a preprocessed data set or a modified version of the preprocessed data set to one or more second deep neural networks trained using one or more sets of biophysical signal data acquired from a plurality of patients flagged as a diagnosis of the presence or absence of coronary artery disease at a coronary artery selected from the group consisting of: left aorta (LMA), proximal left circumflex (Prox LCX), Mid left circumflex (Mid LCX), distal left circumflex (Dist LCX), LPAV, first blunt edge (OM1), second blunt edge (OM2), third blunt edge (OM3), proximal left anterior descending (Prox LAD), Mid left anterior descending (Mid LAD), distal left anterior descending (Dist LAD), LAD 1, LAD 2, proximal right coronary artery (Prox RCA), Mid right coronary artery (Mid RCA), distal right coronary artery (Dist RCA) and acute marginal branch of right posterior descending (AcM R PDA),

wherein an output data set is output via a report and/or display based on the determined values indicative of the presence of a heart disease or condition at the one or more coronary arteries.

22. A method of configuring a convolutional neural network to detect the presence of coronary artery disease or to estimate the location of coronary artery disease in a subject, the method comprising:

generating, by a processor, a plurality of hyper-parameter sets for a template convolutional neural network, wherein each of the plurality of hyper-parameter sets is generated by a random or pseudo-random selection from a set of hyper-parameters, wherein at least one hyper-parameter of the set of hyper-parameters is selected from the group consisting of: batch size, learning rate, convolutional layer, filter size, number of filters in the first convolutional layer, increment of filters in subsequent layers, number of additional dense layers, size of additional dense layers, activation function type, target, expansion rate, and discard;

training, by the processor, the template convolutional neural network for each of a plurality of sets of hyper-parameters, wherein in each evaluation instance, the template convolutional neural network is configured with a set of hyper-parameters of the plurality of sets of hyper-parameters; and

evaluating, by the processor, a trained deep neural network using a first validation dataset for each of a plurality of hyper-parameter sets, wherein each evaluation generates a score, wherein the trained convolutional neural network is subsequently used to diagnose the presence and/or location of coronary artery disease in the subject.

23. The method of claim 22, wherein the assessment of the trained deep neural network comprises generating an accuracy score, a weighted accuracy score, a positive prediction score, a negative prediction score, an F score, a sensitivity score, a specificity score, and/or a diagnostic odds ratio score.

24. A system, comprising:

one or more processors; and

a memory having instructions stored thereon, wherein execution of the instructions by the one or more processors causes the one or more processors to perform the method of any one of claims 1-23.

25. A system, comprising:

a device configured to acquire a broadband phase gradient signal; and

an evaluation system directly or indirectly coupled to the device, the evaluation system comprising:

one or more processors; and

a memory having instructions stored thereon, wherein execution of the instructions by the one or more processors causes the one or more processors to perform the method of any one of claims 1-23.

26. A system, comprising:

a storage area network configured to receive and store acquired wideband phase gradient signal datasets generated from a device configured to acquire wideband phase gradient signals; and

an evaluation system directly or indirectly coupled to the storage area network, the evaluation system comprising:

one or more processors; and

a memory having instructions stored thereon, wherein execution of the instructions by the one or more processors causes the one or more processors to perform the method of any one of claims 1-23.

27. A non-transitory computer-readable medium having instructions stored thereon, wherein execution of the instructions by one or more processors causes the one or more processors to perform the method of any one of claims 1-23.

Technical Field

The present disclosure relates generally to non-invasive methods and systems for characterizing cardiovascular and other physiological systems. More particularly, in one aspect, the present disclosure relates to a non-invasive method for generating phase-space analysis datasets/images from acquired biophysical signals (e.g., cardiac signals, brain/neurological signals, signals associated with other biological systems, etc.) using phase-space data, particularly for predicting and localizing coronary stenosis of the myocardium and characterizing myocardial ischemia, as well as other cardiac and non-cardiac diseases and pathologies.

Background

Ischemic heart disease (also known as cardiac ischemia or myocardial ischemia) is a disease or group of diseases characterized by a reduction in the blood supply to the heart muscle, usually caused by Coronary Artery Disease (CAD). CAD typically occurs when atherosclerosis (hardening or stiffening of the inner wall and the accumulation of plaque therein, often accompanied by abnormal inflammation) occurs in the inner wall of the coronary arteries supplying the myocardium or myocardium. Over time, CAD can also weaken the myocardium and lead to, for example, angina, myocardial infarction (sudden cardiac arrest), heart failure, and cardiac arrhythmias. An arrhythmia is an abnormal heart rhythm that may include any change in the positive sequence of electrical conduction in the heart, which in some cases may lead to cardiac arrest.

The evaluation of CAD can be complex and use many techniques and tools to assess the presence and severity of a disorder. In the case of electrocardiography (i.e., the field of information cardiology in which the electrical activity of the heart is analyzed to obtain information about its structure and function), severe ischemic heart disease can alter the narrowing or blocking of the ventricular conduction properties of the myocardium in the perfusion bed downstream of the coronary arteries. This pathology can manifest itself in different severity levels at different locations of the heart, which makes accurate diagnosis challenging. Furthermore, the conductive properties of the myocardium may vary from person to person, and other factors, such as variability of measurements related to placement of the measurement probe and parasitic losses associated with such probes and their associated components, may also affect the biophysical signals captured during electrophysiological examination of the heart. Furthermore, when the conductive properties of the myocardium are captured as relatively long cardiac phase gradient signals, they may exhibit complex non-linear variability that conventional modeling techniques cannot effectively capture.

Machine learning techniques predict results from a set of input data. For example, machine learning techniques are used to recognize patterns and images, supplement medical diagnosis, and the like. Machine learning techniques rely on a set of features generated using a training dataset (i.e., an observed dataset, the result to be predicted in each observation being known), each feature representing some measurable aspect of the data of the observation band, to generate and adjust one or more predictive models. For example, observed signals (e.g., heart beat signals from multiple subjects) may be analyzed to collect frequency, average, and other statistical information about these signals. Machine learning techniques may use these features to generate and adjust a model that relates the features to one or more conditions, such as some form of cardiovascular disease (CVD), including Coronary Artery Disease (CAD), and then apply the model to data sources for which the results are unknown, such as undiagnosed patients or future patterns, and so forth.

Traditionally, in the context of cardiovascular disease, these features are manually selected from traditional electrocardiograms and combined by a data scientist who works with domain experts.

Disclosure of Invention

The example methods and systems described herein facilitate configuration and training of a neural network (e.g., a deep neural network, a Convolutional Neural Network (CNN), etc.), or all thereof, with a phase gradient biophysical signal dataset (e.g., a wideband phase gradient biophysical signal dataset) to assess and/or classify a disease of a subject. In the context of the heart, the methods and systems described herein facilitate configuration and training of a neural network (e.g., a deep neural network, a Convolutional Neural Network (CNN)), or all thereof, by a phase gradient cardiac signal dataset (e.g., a wideband phase gradient cardiac signal dataset) to assess and/or classify coronary artery disease in a subject. Notably, the exemplary system in this embodiment has been shown to have the following diagnostic capabilities: patients with AUC scores of 0.61 or higher are assessed for global coronary artery disease using a completely non-invasive method of measuring phase gradient biophysical signals of humans on a beat-by-beat basis (also referred to herein as "beat-to-beat"). In some embodiments, the example system is further configured to locate the presence of coronary artery disease in a major coronary artery (e.g., a Right Coronary Artery (RCA), a Left Anterior Descending (LAD) artery, and/or a left circumflex artery (LCX), etc.). In some embodiments, the instantiation system is configured to generate and co-present facies spatial analysis datasets/images with coronary artery disease assessment and localization. Although discussed in the context of cardiac signals, the example methods and systems described herein facilitate configuration and training of neural networks (e.g., deep neural networks, Convolutional Neural Networks (CNNs), etc.), or all thereof, by other biophysical signals (e.g., neural signals, lungs, etc.) to assess and/or classify a disease of a subject or a disease in a particular anatomical structure or organ of a subject.

As used herein, the term "cardiac signal" refers to one or more signals associated with the structure, function and/or activity of the cardiovascular system, including aspects of electrical/electrochemical conduction of the signal, e.g., causing contraction of the myocardium. In some embodiments, the cardiac signal may include electrocardiogram signals, such as those acquired via an Electrocardiogram (ECG) or other form.

As used herein, the term "neural signal" refers to one or more signals associated with the structure, function and/or activity of the central and peripheral nervous systems, including the brain, spinal cord, nerves and their associated neurons and other structures, and the like, and includes the electrical/electrochemical conduction aspects of the signal. In some embodiments, the neural signals may include electroencephalographic signals, such as those acquired through electroencephalography (EEG) or other modalities.

The term "biophysical signal" is not limited to cardiac signals, neurological signals, or photoplethysmographic signals, but encompasses any physiological signal from which information may be obtained. Without limitation by example, biophysical signals may be classified into types or categories that may include: such as electrical signals (e.g., certain heart and nervous system related signals that may be observed, identified, and/or quantified by techniques such as measuring voltage/potential, impedance, resistivity, conductivity, current, etc., in various domains, such as the time and/or frequency domains), magnetic signals, electromagnetic signals, optical signals (e.g., signals that may be observed, identified, and/or quantified by techniques such as reflection, interference, spectroscopy, absorbance, transmittance, visual observation, photoplethysmography, etc.), acoustic signals, chemical signals, mechanical signals (e.g., signals related to fluid flow, pressure, motion, vibration, displacement, strain), thermal signals, and electrochemical signals (e.g., signals that may be associated with the presence of certain analytes such as glucose). In some cases, biophysical signals may be described in the context of the following physiological systems: for example, the respiratory system, circulatory system (cardiovascular, pulmonary), nervous, lymphatic, endocrine, digestive, excretory, muscle, skeletal, renal/urinary/excretory, immune, epidermal/exocrine, and reproductive systems), organ system (e.g., signals that may be unique to the heart and lungs (because they work together)), or biophysical signals are described in the context of tissues (e.g., muscle, fat, nerve, connective tissue, bone), cells, organelles, molecules (e.g., water, proteins, fats, carbohydrates, gases, free radicals, inorganic ions, minerals, acids, and other compounds, elements, and sub-atomic components thereof). Unless otherwise indicated, the term "biophysical signal acquisition" generally refers to any passive or active manner of acquiring a biophysical signal from a physiological system (e.g., a mammalian or non-mammalian organism). Passive and active biophysical signal acquisition generally refers to the observation of natural or induced electrical, magnetic, optical, and/or acoustic emissions of body tissue. Non-limiting examples of passive and active biophysical signal acquisition means include, for example, observing natural radiation of body tissue in voltage/potential, current, magnetic, optical, acoustic and other non-active ways, and in some cases, inducing such radiation. Non-limiting examples of passive and active biophysical signal acquisition means include active means such as ultrasound, radio waves, microwaves, infrared and/or visible light (e.g., for pulse oximetry or photoplethysmography), visible light, ultraviolet light, etc., to interrogate body tissue (not involving ionizing energy or radiation (e.g., X-rays)). Active biophysical signal acquisition may involve excitation-emission spectroscopy (including, for example, excitation-emission fluorescence). Active biophysical signal acquisition may also involve the transmission of ionizing energy or radiation (e.g., X-rays) (also referred to as "ionizing biophysical signals") to body tissue. Passive and active biophysical signal acquisition means may be performed in conjunction with invasive procedures (e.g., by surgery or invasive radiological intervention protocols) or may be performed non-invasively (e.g., by imaging).

As used herein, "photoplethysmographic signal" refers to a signal waveform acquired from an optical sensor that corresponds to a measured change in light absorption of oxygenated and deoxygenated hemoglobin, such as light having wavelengths in the red and infrared spectra. In some embodiments, the photoplethysmography signal comprises a raw signal acquired by a pulse oximeter or a photoplethysmogram (PPG). In some embodiments, the photoplethysmographic signals are acquired from custom or dedicated instruments or circuits (including off-the-shelf devices) configured to acquire such signal waveforms for the purpose of diagnosing a disease or abnormal condition. The photoplethysmography signals typically include red photoplethysmography signals (e.g., electromagnetic signals having primarily wavelengths of about 625 to 740 nanometers in the visible spectrum) and infrared photoplethysmography signals (e.g., electromagnetic signals extending up to about 1 millimeter from the nominal red edge of the visible spectrum), although other spectra (e.g., near infrared, blue, and green) may also be used in different combinations depending on the type and/or mode of PPG employed.

The methods and systems described in various embodiments herein are not so limited and may be used in any context of another physiological system or system, organ, tissue, cell, etc. of a living subject. By way of example only, two biophysical signal types that may be used in a cardiovascular environment include cardiac signals that may be acquired by conventional electrocardiogram (ECG/EKG) devices, bipolar broadband biopotential (cardiac) signals that may be acquired from other devices, such as those described herein, and signals that may be acquired by various volume description techniques, such as photoplethysmography.

In the context of the present disclosure, techniques are described for acquiring and analyzing biophysical signals, in particular for diagnosing the presence, absence, location (where applicable) and/or severity of certain disease states or conditions in, associated with or affecting the cardiovascular (or cardiac) system, including, for example, Pulmonary Hypertension (PH), Coronary Artery Disease (CAD), and heart failure (e.g., left or right heart failure).

Pulmonary hypertension, heart failure, and coronary artery disease are three diseases/conditions associated with the cardiovascular or cardiac system. Pulmonary arterial hypertension (PH) generally refers to hypertension in the arteries of the lungs, which may include a range of conditions. PH often has a complex and multifactorial etiology and has occult clinical episodes of varying severity. PH can develop complications, such as right heart failure, and in many cases is fatal. The World Health Organization (WHO) classifies PH into five groups or 5 types. The first PH group of WHO classification is Pulmonary Arterial Hypertension (PAH). PAH is a chronic disease that is currently incurable, and in particular, it causes the pulmonary artery wall to tighten and stiffen. PAH requires at least cardiac catheterization for diagnosis. PAH is characterized by pulmonary arterial vasculopathy, defined in cardiac catheterization as a mean pulmonary arterial pressure of 25mmHg or higher. One form of pulmonary hypertension is known as idiopathic pulmonary hypertension-PAH that occurs without a clear cause. In addition, a subset of PAHs includes hereditary PAHs, drug and toxin induced PAHs, as well as PAHs associated with other systemic diseases (e.g., connective tissue disease, HIV infection, portal hypertension, and congenital heart disease). PAH includes all causes that result in narrowing of the pulmonary vascular structure. In PAH, the gradual narrowing of the pulmonary artery bed is due to an imbalance of vasoactive mediators (including prostacyclin, nitric oxide and endothelin-1). This can lead to increased right ventricular afterload, right heart failure and premature death. The second PH group of the WHO classification is pulmonary hypertension due to left heart disease. This group of diseases is often characterized by problems on the left side of the heart. Over time, these problems can cause changes in the pulmonary arteries. Specific subgroups include left ventricular systolic dysfunction, left ventricular diastolic dysfunction, valvular disease, and ultimately congenital cardiomyopathy and obstruction due to non-valvular disease. Treatment of the second PH group often focuses on potential problems (e.g., heart valve replacement surgery, various medications, etc.). The third PH group of the WHO classification is large and diverse and is often associated with pulmonary disease or hypoxia. Subgroups include chronic obstructive pulmonary disease, interstitial lung disease, sleep disordered breathing, alveolar hypoventilation, chronic high altitude exposure, and developmental lung disease. The fourth PH group is classified by WHO as a chronic thromboembolic pulmonary hypertension whose blood clots enter or are caused upon lung formation, blocking the flow of blood through the pulmonary arteries. The fifth PH group is classified by the WHO as including subgroups of rare diseases that cause PH, such as hematological diseases, systemic diseases (e.g., sarcoidosis which affects the lungs), metabolic disorders and other diseases. The mechanism of PH in this fifth group is poorly understood.

All forms of PH are difficult to diagnose in routine physical examination, as the most common symptoms of PH (shortness of breath, fatigue, chest pain, edema, palpitations, dizziness) are associated with many other conditions. Blood tests, chest X-ray tests, electrocardiography and echocardiography, pulmonary function tests, exercise tolerance tests, and nuclear scanning are all widely used to help physicians diagnose certain forms of PH. As mentioned above, the "gold standard" for diagnosing PH, and in particular PAH, is cardiac catheterization of the right side of the heart by directly measuring the pressure in the pulmonary artery. If the subject is suspected of having PAH, one of a plurality of surveys may be conducted to confirm the condition, such as electrocardiogram, chest radiograph, and lung function tests, among others. It is commonly seen that electrocardiograms show right heart strain and chest radiographs show evidence of significant pulmonary artery or cardiac hypertrophy. However, a normal electrocardiogram and chest radiograph cannot exclude the diagnosis of PAH. Further testing may be required to confirm diagnosis and to determine etiology and severity. For example, blood tests, exercise tests, and overnight oximetry tests may be performed. In addition, imaging tests may also be performed. Examples of imaging tests include isotope perfusion lung scans, high resolution computed tomography scans, computed tomography lung angiography, and magnetic resonance lung angiography. If these (and possibly other) non-invasive examinations support a diagnosis of PAH, right heart catheterization is often required to confirm the diagnosis by directly measuring lung pressure. It also allows measurement of cardiac output and estimation of left atrial pressure using pulmonary artery wedge pressure. Although non-invasive techniques exist to determine whether PAH is present in a subject, these techniques do not reliably confirm the diagnosis of PAH unless invasive right heart catheterization is performed. Aspects and embodiments of methods and systems for assessing PH are disclosed in commonly-owned U.S. patent application No. 16/429,593, which is incorporated herein by reference in its entirety.

Heart failure affects nearly 600 million people in the united states alone, with over 870,000 people diagnosed with heart failure each year. The term "heart failure" (sometimes referred to as congestive heart failure or CHF) generally refers to a chronic, progressive condition or process in which the heart muscle is unable to pump enough blood to meet the needs of the body, either because the heart muscle is weakened or stiff, or because there is a defect that prevents normal circulation. This can result in, for example, blood and fluid regurgitation to the lungs, edema, fatigue, dizziness, fainting, a rapid and/or irregular heartbeat, dry cough, nausea and shortness of breath. Common causes of heart failure are Coronary Artery Disease (CAD), hypertension, cardiomyopathy, arrhythmia, kidney disease, heart defects, obesity, smoking, and diabetes. Diastolic Heart Failure (DHF), left or left sided heart failure/disease (also referred to as left ventricular heart failure), right or right sided heart failure/disease (also referred to as right ventricular heart failure) and Systolic Heart Failure (SHF) are common types of heart failure.

Left-sided heart failure is further divided into two main types: systolic failure (or heart failure with reduced ejection fraction or reduced left ventricular function) and diastolic failure/dysfunction (or heart failure with preserved ejection fraction or preserved left ventricular function). Procedures and techniques commonly used to determine whether a patient has left-sided heart failure include cardiac catheterization, X-ray, echocardiogram, Electrocardiogram (EKG), electrophysiology studies, radionuclide imaging, and various treadmill tests, including tests that measure peak VO 2. Ejection Fraction (EF) is a measure expressed as a percentage of the amount of blood pumped out by the ventricle on each contraction (left ventricle in the case of left-sided heart failure), most commonly obtained via echocardiography in a non-invasive manner. The normal Left Ventricular Ejection Fraction (LVEF) ranges from about 55% to about 70%.

When systolic failure occurs, the left ventricle cannot contract vigorously to maintain normal systemic blood circulation, thereby depriving the body of the normal blood supply. As the left ventricle pumps harder for compensation, it becomes weaker and thinner. As a result, blood flows back into the organ, causing fluid accumulation in the lungs and/or swelling in other parts of the body. Echocardiography, magnetic resonance imaging, and nuclear medicine scanning (e.g., multi-gated acquisition) are techniques for non-invasively measuring the Ejection Fraction (EF), expressed as a percentage of the left ventricular pump volume relative to its fill volume, to help in diagnosing systolic failure. In particular, Left Ventricular Ejection Fraction (LVEF) values below 55% indicate that the pumping capacity of the heart is lower than normal, and in severe cases can be measured to be below about 35%. In general, when these LVEF values are lower than normal, diagnosis of systolic failure can be made or facilitated.

When diastolic heart failure occurs, the left ventricle becomes stiff or thicker, losing its ability to relax normally, which in turn means that the left lower chamber of the heart is not properly engorged with blood. This reduces the amount of blood pumped out to the body. Over time, this can lead to blood accumulation in the left atrium and then in the lungs, leading to fluid blockage and heart failure symptoms. In this case, the LVEF value tends to remain within the normal range. Thus, other tests (e.g., invasive catheterization) may be used to measure Left Ventricular End Diastolic Pressure (LVEDP) to help diagnose diastolic heart failure as well as other forms of heart failure with normal EF. Typically, LVEDP is measured directly by placing a catheter in the left ventricle or indirectly by placing a catheter in the pulmonary artery to measure pulmonary capillary wedge pressure. By their nature, such catheterization techniques increase the risk of infection and other complications for the patient and are often expensive. Accordingly, a non-invasive method and system for determining or estimating LVEDP in diagnosing the presence or absence and/or severity of diastolic heart failure, as well as myriad other forms of heart failure with normal EF, is desirable. Furthermore, non-invasive methods and systems (not necessarily including the determination or estimation of abnormal LVEDP) for diagnosing the presence or absence and/or severity of diastolic heart failure, as well as myriad other forms of heart failure with normal EF, are desirable. Embodiments of the present disclosure address all of these needs.

Right-sided heart failure typically occurs as a result of left-sided heart failure, when the weak and/or stiff left ventricle loses its ability to effectively pump blood to other parts of the body. As a result, fluid is forced back through the lungs, weakening the right side of the heart, resulting in right-side heart failure. This reflux backs up in the veins, causing fluid in the legs, ankles, gastrointestinal tract and liver to expand. In other cases, certain lung diseases (e.g., chronic obstructive pulmonary disease and pulmonary fibrosis) can lead to right-sided heart failure, even if left-sided cardiac function is normal. Procedures and techniques commonly used to determine whether a patient has left-sided heart failure include blood tests, cardiac CT scans, cardiac catheterization, X-ray, coronary angiography, echocardiography, Electrocardiogram (EKG), myocardial biopsy, lung function studies, and various forms of stress testing, such as treadmill testing.

Pulmonary hypertension is closely associated with heart failure. As mentioned above, PAH (first WHO PH grouping) can lead to increased right ventricular afterload, right heart failure, and premature death. PH caused by left heart failure (second WHO PH group) is considered to be the most common cause of PH.

Ischemic heart disease (also known as cardiac ischemia or myocardial ischemia) and related conditions or pathologies may also be estimated or diagnosed using the techniques disclosed herein. Ischemic heart disease is a disease or group of diseases characterized by a reduced blood supply to the heart muscle, usually caused by Coronary Artery Disease (CAD). CAD is closely related to heart failure, and is the most common cause thereof. CAD typically occurs when atherosclerosis (hardening or stiffening of the inner wall and the accumulation of plaque therein, often accompanied by abnormal inflammation) occurs in the inner wall of the coronary arteries that supply the heart muscle or muscles. Over time, CAD can also weaken heart muscle and cause, for example, angina, myocardial infarction (sudden cardiac arrest), heart failure, and cardiac arrhythmias. An arrhythmia is an abnormal heart rhythm that may include any change in the normal sequence of electrical conduction through the heart, which in some cases may lead to cardiac arrest. The assessment of PH, heart failure, CAD, and other diseases and/or conditions can be complex and use a number of invasive techniques and tools to assess the presence and severity of such conditions. Furthermore, the commonality between the symptoms of these diseases and/or disorders and the fundamental link between the respiratory system and the cardiovascular system, as they work together to supply oxygen to the cells and tissues of the body, suggests a complex physiological interrelationship that can be used to improve the detection and ultimate treatment of such diseases and/or disorders. In such cases, traditional methods of assessing these biophysical signals still present significant challenges in providing healthcare providers with tools for accurately detecting/diagnosing the presence or absence of such diseases and disorders.

For example, in electrocardiography, the field of cardiology in which the electrical activity of the heart is analyzed to obtain information about its structure and function, it has been observed that significant ischemic heart disease can alter the ventricular conduction properties of the perfused bed myocardium downstream of the coronary arteries, causing it to narrow or occlude, and pathology can manifest itself at different locations and at different stages of severity in the heart, making accurate diagnosis challenging. Furthermore, the conductive properties of the myocardium may vary from person to person, and other factors, such as measurement variability associated with the placement of measurement probes and parasitic losses associated with such probes and their associated components, may also affect the biophysical signals captured during cardiac electrophysiology examinations. Furthermore, when the conductive properties of the myocardium are captured as relatively long cardiac phase gradient signals, they may exhibit complex non-linear variability that conventional modeling techniques cannot effectively capture.

In one aspect, a method is disclosed (e.g., using a phase gradient biophysical signal dataset (e.g., a wideband phase gradient biophysical signal dataset, a phase gradient cardiac signal dataset, a wideband phase gradient cardiac signal dataset) to facilitate configuration and training of a neural network (e.g., a deep neural network, a Convolutional Neural Network (CNN), etc.) or all thereof to assess and/or classify coronary artery disease in a subject). The method comprises the following steps: receiving, by a processor, a set of biophysical signal data of a subject acquired from one or more channels of one or more sensors; and pre-processing the biophysical signal data set to generate one or more pre-processed data sets, wherein each pre-processed data set comprises a single isolated complete cardiac cycle (wherein the pre-processed data sets from each acquisition channel are phase synchronized/aligned); and determining, by the processor, a value (e.g., risk/likelihood, binary indication) indicative of the presence or absence of a heart disease or condition (e.g., coronary artery disease, pulmonary hypertension, left heart failure, right heart failure, and left ventricular end diastolic pressure disorder (LVEDP)) by inputting the preprocessed dataset directly to, or all of, one or more neural networks (e.g., one or more deep neural networks, one or more convolutional neural networks, etc.) trained using a set of training biophysical signal datasets acquired from patients: the patient is diagnosed with a heart disease or condition and is tagged with the presence or absence of a heart disease or condition (e.g., where the tagging is based on a genini score or binary value of the disease location in the coronary arteries) (e.g., where the segmented data sets are phase-aligned between the respective biophysical signal data sets of the other acquisition channels), wherein an output data set is output via a report and/or display based on the determined value indicative of the presence of a heart disease or condition (e.g., to assist or for diagnosing the presence or absence of a heart disease or condition in the subject).

In some embodiments the heart disease or condition is coronary artery disease, and wherein the step of determining a value indicative of the presence of a heart disease or condition comprises: inputting (e.g., directly inputting) the preprocessed data set into a set of one or more neural networks (e.g., a set of one or more deep neural networks, a set of one or more convolutional neural networks, etc.), or all thereof, the deep neural network is trained using one or more sets of biophysical signal data acquired from a plurality of patients labeled with a diagnosis of the presence or absence of coronary disease (e.g., significant coronary disease), (e.g., a label for labeling the presence of coronary disease includes a Gensini-based score determined as a combination of a severity-weighted score and a location-weighted score for coronary lesions diagnosed in the myocardium) wherein an output of the one or more neural networks (e.g., an output of the deep neural network, an output of a convolutional neural network), or all thereof, is output as an output data set via the report and/or the display.

In some embodiments, the biophysical signal data sets are acquired from two or more acquisition channels, and the preprocessed data sets from each of the acquisition channels are phase synchronized.

In some embodiments, the step of preprocessing the biophysical signal dataset comprises: segmenting, by the processor, a portion of a biophysical signal dataset associated with a first acquisition channel of the one or more acquisition channels or a normalized dataset derived from the portion of the biophysical signal dataset into one or more first segmented datasets, wherein each of the first segmented datasets comprises a single isolated complete cardiac cycle (e.g., for a per-beat analysis) as a first single isolated complete cardiac cycle, wherein the first single isolated complete cardiac cycle has an associated time window; and segmenting, by the processor, another portion of the biophysical signal data set associated with a second acquisition channel of the one or more acquisition channels or a normalized data set derived from another portion of the biophysical signal data set into one or more second segmented data sets, wherein each of the one or more second segmented data sets comprises a second single isolated complete cardiac cycle, wherein the second single isolated complete cardiac cycle has an associated time window corresponding to the time window of the first single isolated complete cardiac cycle to provide a phase-synchronized data set.

In some embodiments, the label for marking the presence of coronary artery disease comprises a Gensini-based score determined as a combination of a severity-weighted score and a location-weighted score for coronary lesions diagnosed in the myocardium.

In some embodiments, the Gensini-based score is linearized (e.g., via a logarithmic operator).

In some embodiments, the method comprises: determining, by the processor, one or more location values indicative of the presence or absence of a heart disease or condition at a given coronary artery by inputting (e.g., directly inputting) the preprocessed data set or a modified version of the preprocessed data set to one or more second neural networks (e.g., one or more second deep neural networks, one or more second convolutional neural networks), or all thereof, the second deep neural network being trained using one or more sets of biophysical signal data acquired from a plurality of subjects (e.g., a coronary artery disease localization array), the plurality of patients being labeled (binary labeled) as a diagnosis of the presence or absence of a coronary artery disease at a coronary artery selected from the group consisting of: a left aorta (LMA), a proximal left circumflex (Prox LCX), a Mid left circumflex (Mid LCX), a distal left circumflex (Dist LCX), LPAV, a first blunt edge (OM1), a second blunt edge (OM2), a third blunt edge (OM3), a proximal left anterior descending (Prox LAD), a Mid left anterior descending (Mid LAD), a distal left anterior descending (Dist LAD), LAD 1, LAD 2, a proximal right coronary artery (Prox RCA), a Mid right coronary artery (Mid RCA), a distal right coronary artery (Dist RCA) and an acute marginal branch of the right posterior descending (AcM R PDA), wherein the determined one or more location values are output as an output dataset via the report and/or the display.

In some embodiments, the method further comprises: comparing, by the processor, the value indicative of the presence of the cardiac disease or condition (e.g., risk/likelihood, binary indication) to a threshold, wherein the step of determining one or more location values indicative of the presence of the cardiac disease or condition at a given coronary artery is performed based on the comparison (e.g., wherein the value indicative of the presence of the cardiac disease or condition indicates a positive status of the cardiac disease or condition).

In some embodiments, the method further comprises: performing, by the processor, a phase space operation on the received biophysical signal dataset or the preprocessed dataset to generate one or more phase space datasets/images; and outputting, by the processor, one or more generated facies space datasets/images, wherein the one or more generated facies space datasets/images are presented in the report and/or the display simultaneously and/or synchronously with the output dataset.

In some embodiments, the step of preprocessing the biophysical signal dataset to generate one or more preprocessed datasets further comprises a second preprocessing operation selected from the group consisting of: performing a downsampling operation; performing a baseline wander removal operation; and performing a normalization operation (e.g., normalizing the data set to between 0 and 1).

In some embodiments, at least one or all of the one or more neural networks (e.g., one or more deep neural networks, one or more convolutional neural networks, etc.) are configured based on a hyper-parametric search cycle, wherein implementation of the hyper-parametric search cycle comprises: generating, by the processor, a plurality of hyper-parameter sets for a template neural network (e.g., a template deep biblical network, a template convolutional neural network, all templates thereof, etc.), wherein each of the plurality of hyper-parameter sets is generated by random or pseudo-random selection from a set of candidate hyper-parameters, at least one of the candidate hyper-parameter sets being selected from the group consisting of: batch size, learning rate, convolutional layer, filter size, number of filters in the first convolutional layer, increment of filters in subsequent layers, number of additional dense layers, size of additional dense layers, activation function type, target, expansion rate, and discard; training, by the processor, the template neural network for each of a plurality of sets of hyper-parameters, wherein, in each evaluation instance, the template neural network is configured with a set of hyper-parameters of the plurality of sets of hyper-parameters; and training, by the processor, a template neural network for each of a plurality of sets of hyper-parameters, assessing, by the processor, the trained neural network (e.g., the trained deep neural network, the trained convolutional neural network, etc.) or all thereof using a first validation data set, wherein each assessment generates a score (e.g., "area under curve" or AUC score or true AUC score).

In some embodiments, at least one of the one or more neural networks (e.g., one or more deep neural networks, one or more convolutional neural networks, etc.), or all thereof, is configured based on bayesian hyper-parameter optimization.

In some embodiments, the assessment of the trained neural network (e.g., the trained deep neural network, the convolutional neural network, etc.) or all thereof includes generating an accuracy score, a weighted accuracy score, a positive prediction score, a negative prediction score, an F score, a sensitivity score, a specificity score, and/or a diagnostic odds ratio score.

In some embodiments, at least one or all of the one or more second neural networks (e.g., one or more second deep neural networks, one or more second convolutional neural networks, etc.) are configured based on a hyper-parameter search loop (e.g., where at least one hyper-parameter of the set of hyper-parameters used for configuration is selected from the group consisting of a batch size, a learning rate, a convolutional layer, a filter size, a number of filters in a first convolutional layer, an increment amount of filters in a subsequent layer, a stride, a number of additional dense layers, a size of additional dense layers, an activation function type, a maximum pool size, a drop, and a dilation rate).

In some embodiments, the one or more sets of biophysical signal data acquired from a plurality of patients marked with a diagnosis of the presence or absence of coronary artery disease at the coronary arteries are configured as a coronary artery disease localization array, and wherein the localization array comprises a plurality of elements, each element corresponding to a marker indicative of the presence or absence of a heart disease or condition at a given location in the coronary arteries.

In some embodiments, the method further comprises: modifying the value indicative of the presence of the cardiac disease or condition based on one or more additional predictive models, wherein the one or more additional predictive models involve an analysis based on geometric features related to a geometry or topology of the biophysical signal dataset in phase space.

In some embodiments, the method further comprises: merging the value indicative of the presence of a cardiac disease or condition with a second predictive value indicative of the presence of a cardiac disease or condition, wherein the second predictive value indicative of the presence of a cardiac disease or condition is based on one or more additional predictive models, wherein the one or more additional predictive models involve an analysis based on geometric features related to the geometry or topology of the biophysical signal dataset in phase space.

In some embodiments, the geometric feature relating to the geometry or topology of the biophysical signal dataset in phase space comprises at least one of: VDfA-B feature, VDp feature, VR _ VDO _ A-B feature, VDT _ A-B feature, m ADA feature, VRcVDcPA-C feature, AD _ VR _ A feature, tnVDp feature, VDTA-A feature, tnVRp feature, VR _ VDO _ A-A feature, LCXp feature, VDfA-A feature, rA-D feature, rA-C feature, rA-B feature, rA-A feature, and VRc _ VDcPA feature.

In some embodiments, the VDp feature is a quantification of the biophysical signal dataset in a region occupied by the identified ventricular depolarization trajectory in phase space.

In some embodiments, the VDFA feature is a quantification of fiducial points of the biophysical signal dataset in phase space, wherein the fiducial points comprise at least one of: a machine-identified maximum ventricular depolarization, a machine-identified point prior to the maximum ventricular depolarization, and a machine-identified conclusion of ventricular depolarizations.

In one aspect, a method is disclosed, comprising the steps of: receiving, by a processor, a biophysical signal dataset of a subject, wherein the biophysical signal dataset relates to a plurality of phase gradient cardiac signals acquired simultaneously from the subject via a respective number of acquisition channels by at least one electrode; pre-processing a biophysical signal data set from at least one of the acquisition channels to generate one or more pre-processed data sets, wherein each pre-processed data set comprises a single isolated complete cardiac cycle; and determining, by the processor, a value (e.g., risk/likelihood, binary indication) indicative of the presence or absence of a cardiac disease or other condition (e.g., coronary artery disease, pulmonary hypertension, left heart failure, right heart failure, and left ventricular end diastolic pressure disorder (ldedp)) by inputting the preprocessed data set directly into a set of one or more deep neural networks (e.g., a set of one or more deep neural networks, a set of one or more convolutional neural networks, etc.) or all thereof, the one or more deep neural networks being trained using one or more biophysical signal data sets (e.g., one or more phase gradient biophysical signal data sets, one or more phase gradient cardiac signal data sets, etc.) acquired from a plurality of patients or subjects who are tagged with a diagnosis of the presence of a coronary artery disease in the patient or subject (e.g., significant coronary artery disease) (e.g., where the label for marking the presence of coronary artery disease comprises a Gensini-based score determined as a combination of a severity-weighted score and a location-weighted score for a coronary lesion diagnosed in the myocardium, and where the preprocessed data set of a given acquisition channel is segmented in a phase-aligned manner into corresponding biophysical signal data sets of other acquisition channels; wherein the output data set is output via the report and/or the display based on the determined values indicative of the binary presence of the cardiac disease or condition.

In one aspect, a method is disclosed, comprising the steps of: receiving, by a processor, a biophysical signal dataset of a subject, wherein the biophysical signal dataset relates to a plurality of phase cardiac gradient signals acquired simultaneously from the subject via a respective number of acquisition channels by at least one electrode; and determining, by the processor, one or more location values indicative of the presence of a heart disease or condition at one or more coronary arteries by inputting the preprocessed data set or a modified version of the preprocessed data set to one or more second neural networks (e.g., one or more second deep neural networks, one or more second convolutional neural networks, etc.) or all thereof, the second deep neural network being trained using one or more sets of biophysical signal data (coronary artery disease localization arrays) acquired from a plurality of patients or subjects, each of the plurality of patients or subjects being flagged as a diagnosis of the presence or absence of coronary artery disease at a coronary artery or relevant myocardium region selected from the group consisting of: a left aorta (LMA), a proximal left circumflex (Prox LCX), a Mid left circumflex (Mid LCX), a distal left circumflex (Dist LCX), LPAV, a first blunt edge (OM1), a second blunt edge (OM2), a third blunt edge (OM3), a proximal left anterior descending (Prox LAD), a Mid left anterior descending (Mid LAD), a distal left anterior descending (Dist LAD), LAD 1, LAD D2, a proximal right coronary artery (Prox RCA), a Mid right coronary artery (Mid RCA), a distal right coronary artery (Dist RCA), and an acute marginal branch of the right posterior descending (AcM R PDA), wherein the output data set is output via a report and/or a display based on the determined values indicative of the presence of a heart disease or disorder at one or more coronary arteries.

In another aspect, a method of configuring a neural network (e.g., a deep neural network, a convolutional neural network, etc.), or all thereof, to detect the presence of or estimate the location of a coronary artery disease or disorder in a subject is disclosed, the method comprising: generating, by a processor, a plurality of hyper-parameter sets for a template neural network (e.g., a template deep neural network, a template convolutional neural network, etc.), wherein each of the plurality of hyper-parameter sets is generated by a random or pseudo-random selection from a set of hyper-parameters, wherein at least one hyper-parameter of the hyper-parameter sets is selected from the group consisting of: batch size, learning rate, convolutional layer, filter size, number of filters in the first convolutional layer, increment of filters in subsequent layers, number of additional dense layers, size of additional dense layers, activation function type, target, expansion rate, and discard; training, by the processor, the template neural network for each of a plurality of hyper-parameter sets, wherein in each evaluation instance, the template neural network (e.g., a template deep neural network, a template convolutional neural network, etc.) is configured with a hyper-parameter set of the plurality of hyper-parameter sets; and for each of the plurality of sets of hyper-parameters, assessing, by the processor, a trained neural network (e.g., a deep neural network, a convolutional neural network, etc.), using the first validation dataset, wherein each assessment generates a score (e.g., an AUC score or a true AUC score), wherein the trained neural network (e.g., the trained deep neural network, the trained convolutional neural network, etc.) is subsequently used to diagnose the presence and/or location of coronary artery disease in the subject.

In some embodiments, the trained neural network (e.g., a trained deep neural network, a trained convolutional neural network, etc.) includes generating an accuracy score, a weighted accuracy score, a positive prediction score, a negative prediction score, an F score, a sensitivity score, a specificity score, and/or a diagnostic odds ratio score.

In another aspect, a system is disclosed, comprising: one or more processors; and a memory having instructions stored thereon, wherein execution of the instructions by the one or more processors causes the one or more processors to perform a method as described above.

In another aspect, a system is disclosed, comprising: a device for acquiring phase gradient biophysical signals (e.g., a wideband phase gradient biophysical signal dataset, a phase gradient cardiac signal dataset, a wideband phase gradient cardiac signal dataset, etc.); and an evaluation system directly or indirectly coupled to the device, the evaluation system comprising: one or more processors; and a memory having instructions stored thereon, wherein execution of the instructions by the one or more processors causes the one or more processors to perform a method as described above.

In another aspect, a system is disclosed, comprising: a storage area network configured to receive and store acquisition phase gradient signal datasets (e.g., wideband phase gradient biophysical signal datasets, phase gradient cardiac signal datasets, wideband phase gradient cardiac signal datasets, etc.) generated from a device configured to acquire wideband phase gradient signals; and an evaluation system directly or indirectly coupled to the storage area network, the evaluation system comprising: one or more processors; and a memory having instructions stored thereon, wherein execution of the instructions by the one or more processors causes the one or more processors to perform a method as described above.

In another aspect, a non-transitory computer-readable medium having instructions stored thereon, wherein execution of the instructions by one or more processors causes the one or more processors to perform the method as described above is disclosed.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments and together with the description, serve to explain the principles of the methods and systems incorporated herein. The embodiments are better understood from the following detailed description when read in conjunction with the accompanying drawings. The drawings include the following figures:

fig. 1 is a diagram of an illustrative system configured to non-invasively assess the presence or absence of coronary artery disease in a human using a neural network (e.g., a deep neural network, a convolutional neural network, etc.) or all thereof (esembles), according to an exemplary embodiment.

Fig. 2A is a schematic diagram of a system including one or more neural networks (e.g., one or more deep neural networks, one or more convolutional neural networks, etc.), or all thereof, wherein the system is configured to predict the presence of a coronary artery disease or disorder, according to an exemplary embodiment in a cardiovascular environment.

Fig. 2B is a schematic diagram of a system including one or more neural networks (e.g., one or more deep neural networks, one or more convolutional neural networks, etc.), or all thereof, wherein the system is configured to predict the presence/absence of coronary artery disease in the coronary arteries, according to an exemplary embodiment in a cardiovascular environment.

Fig. 2C is a diagram illustrating coronary arteries that may be classified by the neural network for detecting coronary artery disease (e.g., deep neural network, convolutional neural network, etc.) of fig. 2A and 2B, or all thereof, according to an example embodiment.

FIG. 3 is a diagram illustrating the preprocessing operation of FIG. 1 according to an exemplary embodiment.

Fig. 4 is a diagram illustrating beat-to-beat isolation of the biophysical signal dataset of fig. 3 according to an exemplary embodiment in a cardiovascular environment.

Fig. 5 is a diagram illustrating a method for training the neural network of fig. 1 (e.g., a deep neural network, a convolutional neural network, etc.) or all thereof, according to an exemplary embodiment.

FIG. 6 illustrates executable code for building a neural network model (e.g., a deep neural network model, a convolutional neural network model, etc.) from a set of randomly selected hyper-parameters, according to an example embodiment.

Fig. 7 is a diagram illustrating the construction of the neural network model of fig. 6 (e.g., a deep neural network model, a convolutional neural network model, etc.) or all thereof using a development, validation, and gating data set, according to an example embodiment.

Fig. 8A shows a side view of placement of surface electrodes or probes to the chest and back of a subject or patient according to an exemplary embodiment.

Fig. 8B shows a front view of the placement of a surface electrode or probe to the same patient according to an exemplary embodiment.

Fig. 9 is a diagram illustrating a particular pipeline process for generating one or more neural network models (e.g., one or more deep neural network models, one or more convolutional neural network models, etc.) configured to non-invasively assess the presence or absence of a coronary artery disease or condition in a human, according to an example embodiment.

Fig. 10 is a diagram illustrating a process for selecting a neural network model (e.g., a deep neural network model, a convolutional neural network model, etc.) configured to non-invasively assess the presence or absence of a coronary artery disease or disorder in a human, according to an example embodiment.

FIG. 11 illustrates an exemplary computing environment in which embodiments and aspects can be implemented.

Detailed Description

Each feature described herein, and each combination of two or more such features, is included within the scope of the present invention provided that the features included in such a combination are not mutually inconsistent.

Although the present disclosure relates to beneficial assessment of biophysical signals in the diagnosis and treatment of heart-related pathologies and disorders and/or neurological-related pathologies and disorders, such assessment is applicable to the diagnosis and treatment (including surgery, minimally invasive and/or drug treatment) of any pathology or disorder in which biophysical signals are involved in any relevant system of a living body. An example of a cardiac environment is the diagnosis of CAD and its treatment (alone or in combination) by various therapies, such as placement of stents in coronary arteries, performing atherectomy, angioplasty, prescribing medication, and/or prescribing exercise, changing nutrition and other lifestyle. Other heart-related pathologies or conditions that may be diagnosed include, for example, cardiac arrhythmias, congestive heart failure, valve failure, pulmonary hypertension (e.g., pulmonary hypertension due to left heart disease, pulmonary hypertension due to pulmonary disease, pulmonary hypertension due to chronic thrombosis, and pulmonary hypertension due to other diseases (e.g., blood or other diseases)), left heart failure, right heart failure, and abnormal Left Ventricular End Diastolic Pressure (LVEDP), among other heart-related pathologies, conditions, and/or diseases. Non-limiting examples of diagnosable nerve-related diseases, pathologies or conditions include, for example, epilepsy, schizophrenia, Parkinson's disease, Alzheimer's disease (and all other forms of dementia), autism spectrum (including Alzheimer's syndrome), attention deficit hyperactivity disorder, Huntington's disease, muscular dystrophy, depression, bipolar disorder, brain/spinal cord tumors (malignant and benign), movement disorders, cognitive disorders, language disorders, various psychoses, brain/spinal cord/nerve injury, chronic traumatic encephalopathy, cluster headache, migraine, neuropathy (various forms including peripheral neuropathy), phantom limb/pain, chronic fatigue syndrome, acute and/or chronic pain (including back pain, back surgery failure syndrome, etc.), motor dysfunction, anxiety disorders, diseases caused by infection or extrinsic factors (such as Lyme disease, Huntington's disease, muscular dystrophy), Encephalitis, rabies), narcolepsy and other sleep disorders, post-traumatic stress disorder, neurological disorders/effects associated with stroke, aneurysm, hemorrhagic damage, and the like, tinnitus and other hearing related diseases/disorders, and vision related diseases/disorders.

Example System

Fig. 1 is a diagram of an illustrative system 100 configured to assess (e.g., non-invasively) the presence or absence of coronary artery disease in a human using a neural network (e.g., a deep neural network, a convolutional neural network, etc.), according to an exemplary embodiment. As described herein, physiological systems may refer to the cardiovascular system, the pulmonary system, the renal system, the nervous system, and other functional systems and subsystems of the body. In the context of the cardiovascular system, the particular embodiment of the system 100 shown in FIG. 1 facilitates the study of the complex nonlinear system of the heart by examining the states or phases in phase space that such a system may exhibit over multiple cycles.

In fig. 1, a measurement system 102 is a non-invasive embodiment (shown as a "measurement system (biophysical)" 102) that acquires a plurality of biophysical signals 104 (e.g., phase gradient biophysical signals) from a subject 106 via any number of measurement probes 114 (shown as probes 114a, 114b, 114c, 114d, 114e, and 114f) to produce a biophysical data set 108. The biophysical signal dataset 108 includes a plurality of acquired signals (e.g., signals acquired from three different channels) that can be combined together to generate a biophysical signal dataset 108 (e.g., a three-dimensional phase space representation) of a multi-dimensional dataset. The measurement system 102 is configured to transmit the acquired biophysical signal dataset 108, or a dataset derived therefrom or processed therefrom, to a repository (e.g., a storage area network) (not shown) accessible by a non-invasive biophysical signal assessment system 110, for example, via a communication transmission system and/or network, or via a direct connection, for evaluation by an analysis engine that performs a phase-space analysis on deterministic chaotic or quasi-periodic features of the acquired biophysical signal dataset 108 to determine a clinical output 112 (including an assessment and/or estimated physiological features of the presence or absence of a disease of the physiological system under study). In some embodiments, the clinical output includes an assessment of the presence or absence of a disease, disorder, or physiological characteristic of the physiological system under study. In other embodiments, there is no clinical output, but rather information is output that clinicians may use to provide their own clinical assessment of information about the patient whose signals are being assessed.

In some embodiments, the measurement system 102 is configured to acquire, via the biopotential sensing circuit, a biophysical signal that may be based on a biopotential of the body as a biopotential biophysical signal. In a cardiac and/or electrocardiographic environment, the measurement system 102 is configured to capture a cardiac-related biopotential signal or electrophysiological signal of a living subject (e.g., a human) as a biopotential cardiac signal data set. In some embodiments, measurement system 102 is configured to acquire a broadband cardiac phase gradient signal as a biopotential signal or other signal type (e.g., a current signal, an impedance signal, a magnetic signal, an optical signal, an ultrasound or acoustic signal, etc.). The term "wideband" with respect to the acquired signals and their corresponding data sets refers to signals having a frequency range significantly greater than the nyquist sampling rate of the highest dominant frequency of the physiological system of interest. For cardiac signals that typically have a dominant frequency component between about 0.5Hz and about 80Hz, the wideband cardiac phase gradient signal or wideband cardiac biophysical signal includes cardiac frequency information at a frequency selected from the group consisting of: between about 0.1Flz and about 1KHz, between about 0.1Hz and about 2KHz, between about 0.1Hz and about 3KHz, between about 0.1Hz and about 4KHz, between about 0.1Hz and about 5KHz, between about 0.1Hz and about 6KHz, between about 0.1Hz and about 7KHz, between about 0.1Hz and about 8KHz, between about 0.1Hz and about 9KHz, between about 0.1Hz and about 10KHz, and between about 0.1Hz and greater than 10KHz (e.g., 0.1Hz to 50KHz or 0.1Hz to 500 KHz). In addition to capturing the primary frequency components, broadband acquisition also facilitates the capture of other frequencies of interest. Examples of such frequencies of interest may include QRS frequency maps (which may range up to 250Hz), and the like. The term "phase gradient" in relation to the acquired signals and the corresponding data set refers to signals acquired at different vantage points of the body to observe phase information for a set of different events/functions of the physiological system of interest. After signal acquisition, the term "phase gradient" refers to the preservation of phase information through the use of non-distorting signal processing and preprocessing hardware, software, and techniques (e.g., phase linear filters and signal processing operators and/or algorithms).

In a neurological environment, the measurement system 102 is configured to capture a neurologically relevant biopotential or electrophysiological signal of a living body (e.g., a human) as a set of neurological biophysical signal data. In some embodiments, measurement system 102 is configured to acquire a broadband neural phase gradient signal as a biopotential signal or other signal type (e.g., a current signal, an impedance signal, a magnetic signal, ultrasound, an optical signal, ultrasound, or an acoustic signal, etc.). Examples of measurement system 102 are described in U.S. patent application publication nos. 2017/0119272 and 2018/0249960, each of which is incorporated by reference herein in its entirety.

In some embodiments, the measurement system 102 is configured to capture a broadband biopotential biophysical phase gradient signal as an unfiltered electrophysiological signal such that spectral components of the signal are not altered. Indeed, in such embodiments, the broadband biopotential biophysical phase gradient signal is captured, converted, or even analyzed (e.g., prior to digitization) without being filtered (which, if filtered, would affect the phase linearity of the biophysical signal of interest) (e.g., by, for example, hardware circuitry and/or digital signal processing techniques, etc.). In some embodiments, the broadband biopotential biophysical phase gradient signals are captured at a microvolt or sub-microvolt resolution at or significantly below the noise floor of conventional electrocardiography, electroencephalography, and other biophysical signal acquisition instruments. In some embodiments, the broadband biopotential biophysical signals are sampled simultaneously with a time offset or "lag" of less than about 1 microsecond, and in other embodiments, with a time offset or lag of no more than about 10 femtoseconds. Notably, the exemplary system minimizes non-linear distortions (e.g., those that may be introduced via certain filters) in the acquired wideband phase gradient signal so as not to affect the information therein.

Still referring to fig. 1A, the evaluation system 110 includes a preprocessing module 116 configured to receive the acquired biophysical signal dataset 108 in the context of cardiac signals and preprocess it to generate one or more preprocessed datasets 118, each of which has a separate set of isolated complete cardiac cycles as a beat-to-beat cardiac dataset.

The evaluation system 110 includes a first set of one or more neural networks 132a (e.g., one or more deep neural networks, one or more convolutional neural networks, etc.), or all thereof, each neural network in this embodiment being trained using a set of training cardiac signal datasets acquired from a patient or subject diagnosed with a cardiac disease or condition. In some embodiments, as shown in fig. 1, the evaluation system 110 includes a second set of one or more neural networks 132b (e.g., a second set of one or more deep neural networks, a second set of one or more convolutional neural networks, etc.) or all thereof, each neural network in this embodiment being trained using a set of training cardiac signal data sets acquired from a patient diagnosed with a cardiac disease or condition and labeled as having a cardiac disease or condition present/localized and/or absent/not localized in a region of a myocardium or a particular coronary artery (e.g., from a set of coronary arteries). In some embodiments, the one or more neural networks 132a receive the preprocessed data set 118 to train a classifier and/or perform classification on the received input. When used for classification, in some embodiments, the output of the first set of one or more neural networks 132a (e.g., one or more deep neural networks, one or more convolutional neural networks, etc.) or all thereof is a value (134a) indicative of the presence of a cardiac disease or condition, such as a binary value or a risk/likelihood score. In some embodiments, the output of the second set of one or more neural networks 132b (e.g., one or more deep neural networks, one or more convolutional neural networks, etc.), or all thereof, is a value (134b), such as a binary value or a risk/likelihood score, indicative of the presence/location of a cardiac disease or condition at a location in the region of the myocardium and/or coronary arteries. In some embodiments, the outputs 134a and 134b are generated from the same neural network or networks (e.g., 132a or 132b) (e.g., one or more deep neural networks, one or more convolutional neural networks, etc.), or all thereof.

As used herein, the term "neural network" (and Artificial Neural Network (ANN)) refers to a series or framework of machine learning algorithms inspired by biological neural networks that can be instantiated in computing hardware and trained to perform tasks, including learning a set of features generated from a set of training data sets, e.g., to optimize one or more predictive models that can be applied to data sources with unknown results. A neural network with a fully connected layer may define a series of functions parameterized by the weights of the network elements. Deep neural networks are examples of such multi-layered interconnected neural networks that are configured to identify patterns directly from a data set with minimal preprocessing. Examples of deep neural network classes include, but are not limited to, for example, feedforward neural networks, recurrent neural networks (recurrent neural networks), multi-layer perceptrons (MLPs), convolutional neural networks, recurrent neural networks (recurrent neural networks), deep belief networks, convolutional deep belief networks, self-organizing maps, deep boltzmann machines, stacked noise reduction auto-encoders, and the like. Convolutional neural networks ("CNNs") and the like are particularly optimized to recognize patterns directly from multi-dimensional datasets (e.g., images). Examples of popular convolutional neural networks include google lenets, ResNets, resenxts, DenseNets, duralpathnets, and the like, each of which can be applied to predict or estimate the presence or absence of a disease state. In some embodiments, the neural network uses deep learning methods such as CNN to classify the multi-dimensional dataset into one or more positive classes and/or one or more negative classes based on machine-extractable features. As used herein, references to one or more neural networks may include one or more instances of the same type of neural network architecture as well as one or more instances or combined instances of different types of neural network architectures.

Description of neural networks is disclosed at http:// cs231nhttp://cs231n.github.io/conventional-networks/The disclosure herein is incorporated by reference in its entirety.

In fig. 1, in some embodiments, the system 100 includes a healthcare provider portal (shown as "portal" 128) configured to display the output of the neural network (e.g., 134a, 134b) (and other data sets) in or with the phase space analysis report and/or the angiographic equivalent report. The portal 128 (which may be referred to as a physician or clinician portal 128 in some embodiments) is configured to access, retrieve, and/or display or present reports and/or outputs (e.g., 134a, 134b) of the neural network (and other data) from a repository (e.g., a storage area network) for reporting.

In some embodiments, as shown in fig. 1, the healthcare provider portal 128 is configured to display the output (e.g., 134a, 134b) (e.g., a deep neural network, a convolutional neural network, etc.) of the neural network or all thereof in or along with the anatomical mapping report 130a, the coronary artery tree report 130b, and/or the 17-segment report 130 c. The portal 128 can present the data, for example, in real-time (e.g., as web page objects), in electronic documents, and/or in other standardized or non-standardized image visualization/medical data visualization/scientific data visualization formats. The healthcare provider portal 128 and/or repository may comply with patient information and other personal data privacy laws and regulations (e.g., the 1996 U.S. health insurance flow and accountability act and the european union general data protection regulations) and laws related to the marketing of medical devices (e.g., the U.S. federal food and drug act and the european union medical instrument regulations). Further description of an exemplary healthcare provider portal 128 is provided in U.S. patent application publication No. 2018/0078146 entitled "Method and System for Visualization of Heart Tissue at Risk," which is incorporated herein by reference in its entirety. Although in some embodiments, the healthcare provider portal 128 is configured to present patient medical information to healthcare professionals, in other embodiments, the portal 128 may be accessible to and useful for patients, researchers, scholars, and/or other portal users.

In some embodiments, the anatomical mapping report 130a includes one or more depictions of a rotatable (and optionally scalable) three-dimensional anatomical map of the cardiac region of the affected myocardium. In some embodiments, the anatomical mapping report 130a is configured to display and switch between a set of one or more three-dimensional views and/or a set of two-dimensional views of the model having the identified myocardial region. In some embodiments, the coronary tree report 130b includes one or more two-dimensional views of the major coronary arteries. In some embodiments, the 17-segment report 130c includes one or more two-dimensional 17-segment views of the corresponding region of the myocardium. In each report, a value (134b) indicating the presence of a heart disease or condition at a location in the myocardium and a label (134a) indicating the presence of a heart disease may be presented as static and dynamic visualization elements (e.g., by color highlighting of the affected myocardium region and by an animated sequence of highlighting the affected coronary artery region) indicating the predicted occlusion region. In some embodiments, each report includes a textual label indicating the presence or absence of cardiac disease (e.g., the presence of significant coronary artery disease) and a textual label indicating the presence (i.e., location) of cardiac disease in a given coronary artery disease.

In some embodiments, in the context of the cardiovascular system, the healthcare provider portal (and corresponding graphical user interface) is configured to present a summary information visualization identifying myocardial tissue and/or occluded coronary arteries of the myocardium at risk. The user interface may be a graphical user interface ("GUI") with a touch-sensitive screen or a pre-touch-sensitive screen having input capabilities. For example, the user interface may be used to guide diagnosis and treatment of a patient and/or to evaluate a patient in a study. For a given study report, the visualization may include a plurality of depictions: a rotatable three-dimensional anatomical map of the cardiac region of the affected myocardium, a corresponding two-dimensional view of the primary coronary artery, and a corresponding two-dimensional 17-segment view of the primary coronary artery to facilitate interpretation and assessment of structural features of the myocardium to characterize abnormalities of cardiac and cardiovascular function. For a given study report, the visualization may include multiple descriptions of the neural network (e.g., one or more deep neural networks, one or more convolutional neural networks, etc.) or outputs of all thereof (e.g., 134a, 134b), e.g., as text labels indicating the presence or absence of a cardiac disease (e.g., the presence of significant coronary artery disease) and text labels indicating the presence of a cardiac disease (i.e., location) in a given coronary artery disease and/or in a relevant myocardial layer region.

To generate the phase space volume dataset/image, the system as shown in fig. 1 comprises a phase space analysis module 124. In some embodiments, the phase-space analysis module 124 helps to separate deterministic chaos of the physiological system from other types of physiological behavior (as opposed to anatomical behavior) that would be displayed as a functional quantification of the physiological system (e.g., as a phase-space analysis dataset/image). In some embodiments, the phase space analysis module 124 is configured to use a model (e.g., generated from a sparse approximation algorithm (e.g., matching pursuit)) to estimate and/or predict deterministic chaos in the preprocessed biophysical signal dataset 118 (or the acquired biophysical signal dataset 108) as residuals of the preprocessed biophysical signal dataset (e.g., subtracted by the model). To model deterministic chaotic behavior and/or characteristics of a physiological system, the analysis engine of the evaluation system 110 is configured to accurately model the acquired biophysical signal dataset (e.g., to greater than 95% accuracy). In some embodiments, the model is generated according to a modeling algorithm (e.g., a sparse approximation algorithm) with a modeling accuracy of greater than 99%. In some embodiments, the modeling algorithm has an accuracy of greater than 99.9%. In some embodiments, the modeling algorithm has an accuracy of greater than 99.99%. In some embodiments, the modeling algorithm has an accuracy of greater than 99.999%. In some embodiments, the modeling algorithm has an accuracy of greater than 99.9999%. In some embodiments, the modeling algorithm is configured to iteratively and recursively select candidate basis functions to add to the model until a stop condition is reached (e.g., the estimated accuracy value reaches a predefined accuracy value (e.g., X%), the model reaches a maximum allowed number of candidates, and/or the model already contains all available candidates).

Examples of useful phase space concepts and analyses are described in the following documents: U.S. patent application publication No. 2018/0000371 entitled "Non-innovative Method and System for Measuring Myocardial Ischemia, Stenosis Identification, Localization and Fractional Flow Reserve Estimation"; U.S. patent application publication No. 2019/0214137 entitled "Method and System to Assembly Disease Using Phase Space volume Objects" filed on 26.12.2018; and U.S. patent application publication No. 2019/0200893 entitled "Method and System to Assembly Disease Using Phase Space Tomography and Machine Learning", each of which is incorporated herein by reference.

Predicting coronary artery disease using neural networks

Fig. 2A is a diagram of a system 100a including one or more neural networks 232A, 232b (e.g., one or more deep neural networks, one or more convolutional neural networks, etc.), or all thereof, configured to predict the presence of coronary artery disease (e.g., in a patient and/or in the location of a coronary artery, according to an exemplary embodiment in a cardiovascular environment). Other neural networks 132, 132b (e.g., deep neural networks, convolutional neural networks, etc.) or all thereof (as described with respect to fig. 1) may be used as an alternative to 232a, 232 b.

Convolutional neural networks (e.g., google nets, renets, rennexts, DenseNets, DualPathNets) include an architecture that may include one or more input layers, one or more CONV layers (e.g., configured to compute dot products between weights of individual neurons and small connected regions), one or more RELU/ELU layers (e.g., including element-by-element activation functions), one or more POOL layers (e.g., including downsampling operators), and one or more FC (fully connected) layers (e.g., including class score computations), among others.

As shown in fig. 2A, the non-invasive measurement system 102 acquires a plurality of biophysical signals 104 from the subject 106 through a plurality of channels via measurement probes or electrodes 114 (shown as probes 114a, 114b, 114c, 114d, 114e, and 114f) to produce a biophysical data set 108 a.

After acquisition, the evaluation system 110a receives the biophysical data set 108a directly or indirectly from the measurement system 102 via a network (i.e., a communications network) or a data repository including a storage area network. The evaluation system 110a includes a preprocessing module 116a and a set of one or more neural networks 232a, 232b (e.g., one or more deep neural networks, one or more convolutional neural networks, etc.), or all thereof, each trained using a set of training biophysical signal data (e.g., a phase gradient biophysical data set, a wideband phase gradient biophysical signal data set) acquired from a patient diagnosed with a heart disease or condition and tagged for the presence of a heart disease or condition in the patient and/or the presence/absence of a heart disease or condition in a particular coronary artery (e.g., from a set of coronary arteries).

In a cardiovascular environment, the preprocessing module 116a is configured to preprocess (by, for example, a phase-linear preprocessing technique) the biophysical data set 108a from at least one of the acquisition channels to generate one or more preprocessed data sets 118a from each acquisition channel, wherein each preprocessed data set 118a includes a single isolated complete cardiac cycle and is phase-aligned with other corresponding isolated complete cardiac cycles in other channels.

The assessment system 110 determines a value (e.g., risk/likelihood, binary indication) indicative of the presence of a cardiac disease or condition (e.g., coronary artery disease, pulmonary hypertension, left heart failure, right heart failure, Left Ventricular End Diastolic Pressure (LVEDP)) by inputting the preprocessed data set directly to a set of one or more neural networks 232a, 232b (e.g., one or more deep neural networks, one or more convolutional neural networks, etc.) or all thereof, the neural networks being trained using one or more biophysical signal data sets acquired from a plurality of patients who are tagged with a diagnosis of the presence of coronary artery disease (e.g., significant coronary artery disease). In some embodiments, the label that marks the presence of coronary artery disease comprises a Gensini-based score determined as a combination of a severity-weighted score and a location-weighted score for a coronary artery lesion diagnosed in the myocardium of the patient.

In some embodiments, the neural network 232a (e.g., a deep neural network, a convolutional neural network, etc.), or all thereof, receives training data that includes the patient/subject's Gensini score (e.g., a modified Gensini score as described herein).

In some embodiments, the neural network 232b (e.g., a deep neural network, a convolutional neural network, etc.), or all thereof, receives training data that includes a binary array in which each element maps to a coronary artery disease diagnosed at a given coronary artery. In some embodiments, the array binary array comprises disease states mapped to coronary arteries selected from the group consisting of: the left aorta (LMA), the proximal left circumflex (Prox LCX), the Mid left circumflex (Mid LCX), the distal left circumflex (Dist LCX), LPAV, the first blunt edge (OM1), the second blunt edge (OM2), the third blunt edge (OM3), the proximal left anterior descending (Prox LAD), the Mid left anterior descending (Mid LAD), the distal left anterior descending (Dist LAD), LAD 1, LAD 2, the proximal right coronary artery (Prox RCA), the Mid right coronary artery (Mid RCA), the distal right coronary artery (Dist RCA) and the acute marginal branch of the right posterior descending (AcM R PDA). Other coronary arteries may also be included.

Fig. 2B is a diagram of a system 100B including one or more neural networks 232a, 232B (e.g., one or more deep neural networks, one or more convolutional neural networks, etc.), or all thereof, configured to predict the presence/location of coronary artery disease in a coronary artery, according to an exemplary embodiment in a cardiovascular environment. As shown in fig. 2B, the non-invasive measurement system 102 acquires a plurality of biophysical signals 104 from the subject 106 via measurement probes 114 (shown as probes 114a, 114B, 114c, 114d, 114e, and 114f) to produce a biophysical data set 108 c.

After acquisition, the evaluation system 110b receives the biophysical data set 108b directly or indirectly from the measurement system 102 via a network or data repository including, for example, a storage area network.

In this embodiment, the evaluation system 110b includes separate preprocessing modules (shown as preprocessing 116b and 116c) for each set of the one or more neural networks (e.g., deep neural networks, convolutional neural networks, etc.) or all thereof. The pre-processing modules 116b, 116c are configured to pre-process the biophysical data set 108b from at least one of the acquisition channels to generate one or more pre-processed data sets (shown as 118b and 118 c).

Further, as shown in fig. 2B, in this embodiment, the evaluation system 110B is configured with a separate aggregation module (shown as "aggregation" modules 202a, 202B) for each of the groups of one or more neural networks 232a, 232B (e.g., one or more deep neural networks, one or more convolutional neural networks, etc.), or all thereof.

Fig. 2C is a diagram illustrating coronary arteries that may be classified by the neural networks of fig. 2A and 2B (e.g., deep neural networks, convolutional neural networks, etc.) or all thereof according to an exemplary embodiment for detecting coronary artery disease. As shown in fig. 2C, the left main coronary artery supplies blood to the left side of the myocardium and is divided into two branches: left Anterior Descending (LAD) artery and left circumflex artery (LCX). The LAD provides blood to the anterior left side of the heart, while the LCX supplies blood to the posterior and lateral sides of the myocardium. RCA supplies the right atrium, right ventricle, and the base of both ventricles and the back of the septum. The location of Coronary Artery Disease (CAD) is critical as it will help cardiologists to formulate strategies for intervention, drug treatment, or both.

Fig. 3 is a diagram illustrating the preprocessing operation of fig. 1 according to an exemplary embodiment in a cardiovascular environment.

The method 300 includes acquiring (as shown at step 302) the biophysical signal dataset 108, for example, from the measurement system 102 or from a data repository that has received the biophysical dataset from the measurement system 102, as described with respect to fig. 1. In some embodiments, six simultaneously sampled signals are captured from a resting subject as a raw differential channel signal data set (e.g., including channels that may be referred to as "ORTH", "ORTH 2", and "ORTH 3"), where the signals embed inter-lead timing and phase information of the subject-specific acquired signals. Geometric contrast resulting from interference of the phase plane of the depolarization wave with other orthogonal leads may be used, which may facilitate superposition of phase-space information on a three-dimensional representation of the heart. The noiseless subspace further facilitates the observation of the phase of these waves. That is, the phases of the orthogonal leads carry information about the structure and produce geometric contrast in the image. Phase contrast (phase contrast) takes advantage of the fact that different bioelectrical structures have different impedances, so spectral and non-spectral conduction delays and bends the trajectory of the phase space trajectory through the heart by different amounts. In terms of cardiovascular environment, these small changes in trajectory can be normalized and beat-to-beat quantified and corrected for abnormal or poor lead placement, and the normalized phase-space integral can be mapped to a geometric grid for visualization.

In some embodiments, the non-invasive measurement system 102 is configured to sample a biophysical signal (e.g., a bipolar biopotential signal) at a sampling rate of approximately greater than 1kHz (e.g., 8kHz) for a duration of between about 30 to about 1400 seconds, such as about 210 seconds, for each of three differential channels placed orthogonally on the subject. Other durations and sampling rates may also be used.

In some embodiments, the evaluation system 110 then removes the baseline from the acquired raw signals at step 304. In some embodiments, the baseline wander removal operation is implemented as a phase-linear second-order high-pass filter (e.g., a second-order forward-backward high-pass filter with a 0.67Hz cutoff frequency). The forward and reverse operations ensure that the resulting preprocessed biophysical signal data set is phase linear. Other phase linear operations may be used, such as based on wavelet filters, etc.

In other embodiments, a multi-stage moving average filter (median filter, e.g., of the order of 1500 milliseconds, smoothed with a 1Hz low pass filter) is used to extract a bias signal from each input raw difference channel signal. The bias is then removed from the signal by subtracting the estimate of the signal from the maximum of the probability density calculated using the kernel smoothing function.

In some embodiments, the signal runs through a signal quality test, where the correlation output is a time index (of sufficient quality) of the signal suitable for analysis. An example of Signal Quality testing is described in U.S. provisional application No./_____________, entitled "Method and System for Automated Quantification of Signal Quality", filed concurrently with this application (attorney docket No. 10321-.

In some embodiments, the evaluation system 110 downsamples the input signal or the pre-processed signal (e.g., to 1kHz) at step 306. In some embodiments, the down-sampling operation is an averaging operator or a decimation operator.

In some embodiments, the method further comprises normalizing the input acquired biophysical signal dataset 108 or the pre-processed signal 118. Similar types of downsampling, baseline wander, and/or normalization operations may be applied to other biophysical signal datasets.

In some embodiments, the method comprises using only a portion of the acquired biophysical signal dataset, e.g. after a predefined time or dataset shift (e.g. after the first 31 seconds). It is observed that in some embodiments, such operations may minimize and/or reduce motion artifacts (and thus improve signal quality) that may be introduced by movement of the subject at the beginning of the measurement acquisition. It has also been observed that such manipulation can minimize and/or reduce distortions in the measurements (and thus improve signal quality) that may be caused by probe placement and contact points, and it is generally observed that such distortions are reduced as the probe is placed during measurement acquisition. Other temporal or dataset migration techniques may be used, for example, techniques based on the quantification of noise in the acquired biophysical dataset, which may be the result of or related to: a biophysical signal acquisition protocol (instructions), the type of probe or electrode used, and the type and/or configuration of the component (e.g., cable used to transmit signals, biophysical signal measurement system, biophysical signal acquisition space/environment, proximity to other medical instruments, etc.).

In some embodiments, the evaluation system 110 extracts a plurality of "clean" sub-signals from the acquired biophysical signal dataset (or other intermediate signals, e.g., as discussed herein) on a beat-to-beat basis at step 308.

Fig. 4 is a diagram illustrating beat-to-beat isolation of fig. 3 according to an exemplary embodiment. As shown in fig. 4, the evaluation system 110 detects a maximum peak (shown as 402a, 402b, 402c) in the acquired biophysical signal dataset 108 (or an intermediate dataset derived from the biophysical signal dataset, e.g. a down-sampled signal dataset), and then isolates each beat as a dataset defined in a fixed window (shown as 404a, 404b, 404c) located around the maximum peak (108a, 108b, 108c) and with the peak located at the center of the window. In some embodiments, the evaluation system 110a uses the Pan-times Algorithm in "a Real-Time QRS Detection Algorithm" by Pan and times, published in IEEE trans biological eng. 32, vol 3 (3 1985), which is incorporated herein by reference in its entirety, to detect peaks (402a, 402b, 402c) in the downsampled signal dataset (108a, 108b, 108 c). In some embodiments, the evaluation system 110 produces a fixed window of approximately 0.75 seconds, which corresponds to a heart rate of 80 beats per second. Other window sizes and centering techniques may be used.

In some embodiments, to preserve phase gradient information in the acquired biophysical dataset (or an intermediate dataset processed in an assessment analysis), the assessment system 110 applies the same time window (e.g., 404a, 404b, 404c) as obtained in peak detection of a first channel (e.g., ORTH1) to extract heartbeats from one or more other channels (e.g., ORTH2 and ORTH 3). As shown, evaluation system 110 generates a first beat-to-beat segment (406a) from channel "1" (also referred to as channel "ORTH 1") that is phase-aligned with beat-to-beat segment (406b) from channel "2" (also referred to as channel "ORTH 2) and beat-to-beat segment (406c) from channel" 3 "(also referred to as channel" ORTH3), which may be used together as one input to a convolutional neural network. The second set of inputs is shown as a beat-to-beat segment from channel "1" (408a), a beat-to-beat segment from channel "2" (408b), and a beat-to-beat segment from channel "3" (408 c). The third set of inputs is shown as a beat-to-beat segment from channel "1" (410a), a beat-to-beat segment from channel "2" (410b), and a beat-to-beat segment from channel "3" (410 c). In practice, the output of the pre-processing module 116, which is a set of data segments (comprising a single complete cardiac cycle) from a phase-aligned time window (e.g., a 0.75 second window) for each acquisition channel or a portion of the acquisition channels, will be provided as input to one or more neural networks (e.g., a deep neural network, such as a convolutional neural network, etc.) or all thereof. In some embodiments, all of the data segments extracted from the pre-processing module 116 are provided as input to the neural network 132 (e.g., a deep neural network, a convolutional neural network, etc.) or all thereof (e.g., for training or analysis). In other embodiments, the data segments extracted from some but not all acquisition channels are provided to the neural network 132 (e.g., a deep neural network, a convolutional neural network, etc.) or all of it (e.g., for training or analysis). In other embodiments, data segments extracted from some, but not all, of the given acquisition channels are provided to the neural network 132 (e.g., a deep neural network, a convolutional neural network, etc.) or all thereof.

Referring back to fig. 3, after extracting the beat-to-beat segments, evaluation system 110 is configured to normalize each of the beat-to-beat segments as an input to a neural network (e.g., a deep neural network, a convolutional neural network, etc.) or all thereof at step 310. In some embodiments, the evaluation system 110 is configured to evaluate the condition according to

Equation 1 scales each heartbeat on each channel:

as shown, the system divides each channel by its maximum absolute value in the window to provide a data set bounded by a range of-1 to + 1. The evaluation system 110 then adds 1 to the result and divides by 2 to provide a signal in the range of 0 to 1. The evaluation system 110 then provides a signal data set with a mean of 0.5 between 0 and 1 by subtracting the result from the mean of the windowed data set and then adding 0.5. The same operation is performed for the other channels to provide similar ranges and means for all training examples, thereby helping the network to learn and generalize better. Other normalization operations may be used.

Still referring to fig. 3, evaluation system 110 is configured to then input the normalized beat-to-beat data set to a neural network 132 (e.g., a deep neural network, a convolutional neural network, etc.) or all thereof. At step 312, the system may apply the normalized beat-to-beat data set to train the neural network 132 (e.g., a deep neural network, a convolutional neural network, etc.), or all thereof. The system may alternatively apply the normalized beat-to-beat dataset to classify the beat-to-beat dataset by a trained neural network (e.g., a trained deep neural network, a trained convolutional neural network, etc.) or all of the trained to predict the presence or absence of a disease state or disorder (e.g., the presence or absence of coronary artery disease or other disorder) and/or the presence/location of a disease or disorder in the coronary arteries in step 314.

In some embodiments, the evaluation system 110 is configured with more than one neural network 132a, 132b (e.g., a deep neural network, a convolutional neural network, etc.), or all thereof. Each or all of the neural networks 132A, 132B (e.g., deep neural networks, convolutional neural networks, etc.) may receive the normalized beat-to-beat data set and generate a combined set of predictors (e.g., via aggregation operators 202 as shown in fig. 2A; or via operators 202A, 202B as shown in fig. 2B).

Method for optimizing/training convolutional neural network

Fig. 5 is a diagram illustrating a method 500 of training a neural network (e.g., a deep neural network, a convolutional neural network, etc.), (e.g., 132a, 132b, 232a, 232b), or all, according to an example embodiment. In fig. 5, the normalized data set provided by the preprocessing process of fig. 3 is provided as input to the training phase. In the training phase, a set of randomly generated neural network configurations (e.g., generated deep neural network configurations, generated convolutional neural network configurations, etc.) or all thereof are trained (e.g., phase gradient biophysical signal data, wideband phase gradient biophysical signal data) using a developed set of biophysical signal data, which is normalized and preprocessed according to the steps discussed with respect to fig. 3 and evaluated using an identified set of phase gradient biophysical signal data sets (evaluation set). In some embodiments, the evaluation system 110 is configured to use the gennini-based score as a portion of an input to a randomly generated neural network (e.g., a randomly generated deep neural network, a randomly generated convolutional neural network, etc.) or all thereof.

In some embodiments, the evaluation system 110 assigns a single Gensini-based score to the subject (i.e., to the acquired wideband phase gradient dataset, and to a normalized dataset derived from the acquired wideband phase gradient dataset) that reflects the overall burden on the myocardium caused by coronary lesions that are localized and quantified by coronary angiography. In some embodiments, the Gensini-based score is based on equation 2 described in "coronar aricerography" of goffredox g.gensini, published in Futura pub.co (1975), which is incorporated herein by reference in its entirety:

Gensini_score＝∑_iSeverity_ixLocation_i(equation 2)

In equation 2, i is the number of identified coronary lesions, severity_iIs a severity weight {1, 2, 4, 8, 16, or 32} of the values of 25%, 50%, 70%, 90%, 99%, and 100% of the diameter reduction for each assessment of coronary lesion i, and location_iIs the position weight {5.. 0.5}, which is the position, from which pair it is for a positionOther positions, the relative effect on the whole coronary circulation. In fact, if a lesion is located further upstream in the coronary circulation path (e.g., near the aorta), then the lesion has a greater effect on the circulation than a lesion located very low down in the circulation (distal). The location score in the example ranges from 5 to 0.5, with a value of 5.0 assigned to the location with the greatest effect on circulation and a value of 0.5 assigned to the location with the least effect (e.g., according to the Gensini metric). Other scoring values may be used.

In some embodiments, the evaluation system 110 assigns a modified Gensini-based score to the subject (i.e., to the acquired phase gradient biophysical data set, the acquired wideband phase gradient biophysical signal data set) or a normalized data set derived therefrom, which reflects the burden caused by worst coronary lesions located and quantified by coronary angiography, according to equation 3:

Gensini_score＝select_max(Severity_ixLocation_i) (equation 3)

In fact, as provided in equation 3, the system only considers the worst case lesion, i.e., the lesion with the maximum value of the severity weight multiplied by the location weight.

Still referring to fig. 5, for each data preparation and learning step, referred to as an epoch (epoch), the assessment system 110 traverses all training data sets and calculates an AUC score from the validation data set. Because a single phase gradient biophysical data set (e.g., a wideband phase gradient biophysical signal data set) can be partitioned into multiple windowed data sets, in some embodiments, the system is configured to combine all or a majority of the predictions of heartbeats of multiple patients by an averaging operator to provide a combined AUC score. Evaluation system 110 may perform the learning and assessment steps in a loop that may run across multiple machines simultaneously without synchronization.

In some embodiments, the evaluation system 110 is configured to sample the data sets in a hierarchical manner so as to have similar ratios of CAD positive and CAD negative data sets in both the training data set and the confirmation set.

Still referring to fig. 5, the evaluation system 110 is configured to generate a random hyper-parameter set and a neural network architecture (e.g., a deep neural network architecture, a convolutional neural network architecture, etc.) from a set of candidates, as provided in table 1, which shows an exemplary hyper-parameter search space for beat-to-beat, neural network-based analysis (e.g., a deep neural network-based analysis, a CNN-based analysis, etc.).

TABLE 1

In practice, the evaluation system 110 is configured to generate a plurality of hyper-parameter sets for, or all of, a template neural network (e.g., a template deep neural network, a template convolutional neural network, etc.), where each of the plurality of hyper-parameter sets is generated by random or pseudorandom selection from a set of hyper-parameters (e.g., batch size, learning rate, convolutional layers, filter size, number of filters in a first convolutional layer, addition of filters in subsequent layers, number of additional dense layers, size of additional dense layers, activation function type, target, dilation rate, and discard). In some embodiments, the evaluation system 110 is configured to optimize a neural network (e.g., a deep neural network, a convolutional neural network, etc.) or all thereof through bayesian hyper-parametric optimization.

Table 2 shows a set of hyper-parametric search space classes and candidate values for neural network based coronary artery disease localization analysis (e.g., deep neural network based coronary artery disease localization analysis, CNN based coronary artery disease localization analysis). Other hyperparametric search space classes and corresponding candidate values within the spirit and equivalents of tables 1 and 2 may be employed. In tables 1 and 2, in some embodiments, the evaluation system 110 uses a single element from the set defined in "{ }" and the evaluation system 110 uses values in the range "[ ]".

TABLE 2

FIG. 6 illustrates executable code for building a neural network model (e.g., a deep neural network model, a convolutional neural network model, etc.) from a randomly selected hyper-parameter set in accordance with an illustrative embodiment. The executable code of fig. 6 is configured to run in a Keras open source neural network library and is shown as Python. Examples of Random search for hyper-parameter optimization based on Random search for hyper-parameter optimization are described in Journal of Machine Learning Research 13, page 281-305 (2 months 2012), published by Bergstra and Yoshua, which is hereby incorporated by reference in its entirety.

Referring to FIG. 5, the evaluation system 110 trains a set of neural network models (e.g., deep neural networks, convolutional neural networks, etc.) or all thereof at step 504 over several time periods, where each time period includes one traversal of the entire training set. At the end of each session, the evaluation system 110 calculates a set of training and validation AUC scores. In some embodiments, the evaluation system 110 selects the minimum of the calculated AUC as the fraction of the time period as the worst case performance to underestimate, rather than overestimate, the predicted performance of the neural network model under study (e.g., deep neural network, convolutional neural network), or a collection thereof.

In practice, the evaluation system 110 is configured to (i) train a template neural network (e.g., a template deep neural network, a template convolutional neural network, etc.) for each of a plurality of sets of hyper-parameters, wherein in an instance of each assessment, the template neural network (e.g., the template deep neural network, the template convolutional neural network, etc.) is configured with a set of hyper-parameters from the plurality of sets of hyper-parameters, and (ii) assess the trained neural network (e.g., the deep neural network, the convolutional neural network, etc.) or all thereof using the first set of validation data for each of the plurality of sets of hyper-parameters, wherein each assessment generates an AUC score (e.g., a true AUC score). In some embodiments, the assessment of the trained neural network 116 (e.g., the trained deep neural network, the trained convolutional neural network, etc.) or all thereof may include generating one or more of an accuracy score, a weighted accuracy score, a positive prediction score, a negative prediction score, an F score, a sensitivity score, a specificity score, and/or a diagnostic odds ratio (diagnostic odds ratio) score. To this end, the evaluation system 110 is configured to determine a value (e.g., risk/likelihood, binary indication) indicative of the presence of a cardiac disease or condition (e.g., coronary artery disease, pulmonary hypertension, left heart failure, right heart failure, abnormal Left Ventricular End Diastolic Pressure (LVEDP)) by inputting the preprocessed data set directly into a set of one or more neural networks (e.g., a set of one or more deep neural networks, a set of one or more convolutional neural networks, etc.) or all thereof, the neural networks being trained using one or more biophysical signal data sets acquired from a plurality of patients who are tagged with a diagnosis of the presence of significant coronary artery disease, wherein the tag comprises a Gensini-based score determined to be in a region of a myocardial layer or a particular coronary artery (e.g., from a set of coronary arteries) a combination of a severity-weighted score and a location-weighted score for the coronary lesion diagnosed.

The evaluation system 110 is configured to determine whether to stop the execution of the training and search cycle for each period at step 506. In some embodiments, the criteria for stopping include whether a predefined number of periods (e.g., 10) have been performed in which no higher score is observed. At the end of the run, the best scoring model is saved along with the selected parameters and various outputs (e.g., predictions for the validation set). The algorithm then runs the next time using the new, different parameters and CNN architecture.

In some embodiments, prior to applying the data set to the training operation, the evaluation system 110 is configured to assess and reject signals having excessive power line noise, high frequency noise, and/or period variability noise. In some embodiments, evaluation system 110 is configured to perform a signal quality test to determine whether the wideband phase gradient signal has sufficient signal quality for subsequent analysis.

True AUC score

In some embodiments, to provide an improved assessment of the predictive (i.e., classification) algorithm, the assessment system 110 is configured to consider the balance of the cost of the error and the target class in determining the AUC score. Factors that may be used include known statistics, performance goals, and metrics (e.g., cost of false positives, cost of false negatives).

For example, if the system predicts that the patient has coronary artery disease, the subject may be subjected to further investigation — thus, the cost of false positives is "low" in terms of patient safety and health, as further investigation will confirm whether the disease is present. On the other hand, if the system predicts that the patient is not ill and that he or she is actually ill, then the subject may not take further action with respect to the diagnosis, and thus the cost of false negatives is "high" because the safety and health of the patient may be compromised. In fact, false negatives can cost patients and clinical teams more than false positives (again, in terms of patient safety and health), so instances of false negatives should be assigned more weight in this case (e.g., the cost of false negatives is twice the cost of false positives).

In some embodiments, the system is configured to generate a modified Receiver Operating Characteristic (ROC) graph that does not include AUC (region under ROC curve) measurements. The AUC curve generally assumes interest in all possible points on the classifier ROC curve.

In some embodiments, instead of an AUC measurement, the system is configured to use a 1:1 cost. In such embodiments, the system may consider an increase in false positives to be acceptable for each increase in true positives (or decrease in false negatives). For example, starting from the bottom left corner of the ROC graph, the system may increment one true positive and one false positive by "1" (along the y-axis) and increase the false positive rate (along the x-axis) by a fraction of "1" (e.g., 1/15) due to class imbalance. In fact, by issuing a line with a slope from the lower left corner of the ROC graph, the system can maintain the desired balance of 1:1 costs.

In some embodiments, instead of AUC measurements, the system is configured to use a 1:2 cost ratio for true positives versus false positives. In such an embodiment, the system may maintain a slope of 7.5 (e.g., as a boundary to maintain a 1:2 cost ratio). The combined line may be referred to as a "cost-sorted" ratio line. In practice, a suitable classifier can be considered to have an ROC curve with points at or above the ratio line classified by cost. Other ratio values, such as statistics, performance goals, and metrics (e.g., false positive costs, false negative costs) may be set based on the above factors.

Developing, validating and gating data sets

FIG. 7 is a diagram illustrating the construction of the convolutional neural network model of FIG. 6 by developing, validating, and gating a data set, according to an exemplary embodiment. As shown in fig. 7, a first set of phase gradient datasets (e.g., a wideband phase gradient biophysical dataset) is used for a training dataset (shown as "training dataset" 702); a second set of phase gradient datasets (e.g., a wideband phase gradient biophysical dataset) is used for the validation dataset (shown as "validation dataset" 704); and a third set of phase gradient data sets (e.g., a wideband phase gradient biophysical data set) is used for gating data sets (shown as "gating data set" 706). The data sets 702, 704, 706 provide different levels of testing. In some embodiments, the training data set is continuously tested to optimize and train a neural network (e.g., a deep neural network, a convolutional neural network, etc.). The validation data set is a reserved set (withhold set) that is only occasionally assessed to confirm the performance of the trained neural network (e.g., a trained deep neural network, a trained convolutional neural network, etc.). The gating data set is used for gating, i.e., moving a trained neural network (e.g., a trained deep neural network, a trained convolutional neural network, etc.) into a final/locked configuration. For example, the strobe may be evaluated only once or twice.

In some embodiments, each phase gradient data set (702, 704, and 706) is assessed by coronary angiography to locate and quantify coronary lesions in the subject. Each of the phase gradient data sets (702, 704, and 706) is preprocessed as described with respect to fig. 3 to produce beat-to-beat cardiac data sets.

Wide band phase gradient cardiac biophysical data set

Fig. 8A and 8B are diagrams illustrating an example of placing surface electrodes, such as probes 114a-114f, at the chest and back of a patient or subject to acquire biopotential signals associated with a cardiac signal data set, according to an exemplary embodiment. FIG. 8A shows a side view of the placement of the surface electrodes 114a-114g to the chest and back of a patient according to an exemplary embodiment. FIG. 8B illustrates a front view of the placement of the surface electrodes 106a-106g to the chest and back of a patient according to an exemplary embodiment. As shown, the surface electrode is located at (i) a first position corresponding to the 5 th intercostal space near the subject's right anterior axillary line; (ii) a second position near the left anterior axillary line corresponding to the 5 th intercostal space; (iii) a third position corresponding to a first intercostal space near the left sternal edge; (iv) a fourth position proximate the left sternal border below the patient's sternum and outside the xiphoid process; (v) a fifth position corresponding to a third intercostal space adjacent the left sternal boundary; (vi) a sixth position adjacent the back, opposite the fifth position, to the left of the patient's spine; (viii) a seventh position, near the upper right quadrant of the patient, corresponds to a second intercostal space along the left axillary line.

Results of the experiment

CADLAD study. A "coronary artery disease-learning algorithm development" (caddad) study was conducted, which involved two distinct phases to support the development and testing of machine learning algorithms.

In the first phase of the CADLAD study, paired clinical data was used to guide the design and development of the machine learning phase of preprocessing, feature extraction and development. That is, the clinical study data collected was divided into three cohorts: training queue (50%), validation queue (25%) and validation queue (25%). Similar to the above-described steps of processing signals from a patient for analysis, each acquired data set is first preprocessed to clean up and normalize the data. After the pre-processing process, a set of features is extracted from the signal, where each set of features is paired with a representation of the true condition — e.g., a binary classification with or without significant CAD present or a scored classification with significant CAD present in a given coronary artery. The final output of phase 1 is a fixed algorithm embedded in the measurement system.

In phase 2 of the CADLAD study, machine learning algorithms are used to provide a determination of significant CAD for a previously untested pool of clinical data (i.e., a validation dataset). The criteria for disease are determined according to criteria defined in the american society for cardiology (ACC) clinical guidelines, in particular an angiographic stenosis of more than 70% or a streamlined blood flow fraction of less than 0.80.

In another aspect of the CADLAD study, an evaluation system was developed that automatically and iteratively explores combinations of features in various functional permutations with the goal of finding those combinations that can be predicted based on successful matching of features. To avoid overfitting the solution of the training data, the validation set is used as a comparator. Once the candidate predictors are developed, they are manually applied to the validation data set to evaluate predictor performance against data that is not used at all to generate the predictor.

A beat-to-beat convolutional neural network.Experiments conducted with data collected from the CADLAD study showed that the exemplified systems (e.g., 110a, 110b) can detect significant Coronary Artery Disease (CAD) through a neural network (e.g., Convolutional Neural Network (CNN)) trained using beat-to-beat segmentation data from a broadband phase gradient biopotential signal dataset. The broadband phase gradient biopotential dataset is only preprocessed to remove baseline drift, normalize the data range, and isolate the acquired data on a beat-by-beat basis for beat analysis.

While machine learning has been shown to be useful for diagnosing irregular heart rhythms (i.e., arrhythmias) from ECG recordings, which is standard of care for diagnosing such conditions, the criteria for diagnosing coronary artery disease typically include invasive angiographic examinations involving cardiac catheterization. The exemplary system beneficially predicts the presence or absence of coronary artery disease by using only non-invasive measurements of the body's biophysical signals.

Method for generating B2B CNN.To generate a convolutional neural network for experimentation, a training system was developed and used to assess a large number of potential architectures and hyper-parameters through random searches.

FIG. 9 is a diagram illustrating a detailed pipeline process for generating one or more convolutional neural network models configured to non-invasively assess the presence or absence of coronary artery disease in a human, according to an exemplary embodiment.

In the experiment shown in fig. 9, the system in this configuration retrieves raw phase signal data of the patient at 8kHz from an acquisition measurement device (e.g., a phase signal recorder or PSR) at step 902. The system removes the unwanted baseline signals from the acquired raw signals to generate a centralized data set at step 904. The system uses a second order forward-backward filter configured to not introduce any phase distortion (no phase response). The filter is configured with an effective high-pass cut-off frequency of 0.8 Hz. Separately, the system assesses the signal quality of the collected signals and rejects any collected signals that fail the test from subsequent analysis.

Then, at step 906, the system downsamples the centered dataset from the acquired 8KHz sampling rate to 1KHz using an averaging operator to generate a downsampled centered dataset.

After the downsampling operation 906, the system extracts a set of heartbeat segment data from the downsampled centralized data set at step 908, each heartbeat segment data including a single isolated complete cardiac cycle. The system uses the Pan-Tompkins Algorithm described in "AReal-Time QRS Detection Algorithm" published by Pan et al in IEEE Trans.Biomed.Eng. Vol.32, No. 3 (3 months 1985) for detecting peaks and isolating each complete cardiac cycle for each acquisition channel. In this experiment, the output was a fixed window data set of approximately 0.75 seconds, centered at the highest amplitude point, encompassing the complete cardiac cycle, to provide alignment between all of the heartbeat segment data sets. The same process is used to extract a set of cardiac cycle data from each acquisition channel. During the experiment, only data collected from two of the measurement channels (i.e., data from channels ORTH1 and ORTH3) were used in the analysis because of the periodic variability noise observed in one of the three collection channels, but in other embodiments data for all three measurement channels or any two or one channel may be used in the analysis.

The system then normalizes the numerical range of each extracted heartbeat data set at step 910. The system normalizes each heartbeat segment data for each channel by dividing the data set by the determined maximum absolute value of the data for a given window, thereby confining the data to a range between-1 and + 1. The system then scales down to +0.5 to-0.5 and adds an offset of 0.5 to adjust the range to 0.0 to 1.0. Thus, in this experiment, the mean value of each heartbeat segment data set for each channel was 0.5, ranging between 0.0 and 1.0. Through the normalization and alignment operations, the input of a given CNN receives similar ranges and means of all training data sets, resulting in a stronger classifier. Normalization also makes the signal unitless.

The data set collected from the CADLAD study was partitioned into a development pool, a validation pool, and a gating pool at step 912 of the experiment, where the development pool and validation pool were used for training and initial validation, and the validation pool and gating pool were used for validation and gating.

The system was trained and validated using the following data sets: data sets acquired from 730 patients using a first generation phase space recorder (versions 1.0 and 1.1) configured with unipolar broadband phase gradient voltage capture for training and validation; and a data set acquired from 334 patients using a second generation phase space recorder (version 1.2) configured with bipolar broadband phase gradient voltage capture. It is observed that using data sets from different acquisition systems improves the performance of the predictor compared to using data sets from a single hardware type. The system selects the assessed CNN model with AUC ≧ 0.57. The system also used a second validation dataset comprising data collected from 164 patients using a second generation facies space recorder (version 1.2). The system also used a third gated data set comprising data acquired from 243 patients using a second generation phase space recorder (version 1.2).

Experiments were performed using Python 3. The packages used included NumPy, Pandas, SciKit-Learn and Keras, and TensorFlow was used for back-end analysis of the neural network. All development and experiments were performed on Amazon Web Services (AWS) servers.

At step 914, the system rejects the acquired biophysical signal dataset with excessive power line interference noise, excessive high frequency noise, and excessive cycle variability noise from being used as a training, validation, or gating dataset. Once all of these preprocessing steps are completed, the model search cycle begins. This cycle can run indefinitely across multiple machines simultaneously without synchronization. A typical run on 4 p2.xlarge AWS servers will be run for 60 hours in the experiment, which time was found to be sufficient to generate a model that satisfies the confirmed AUC of 0.6.

The system generates a random validation set of 250 signals from the development set for each run through the model search loop at step 916. Hierarchical sampling is used to have the same ratio of CAD positive dataset and CAD negative dataset in the training and confirmation sets by using the StratifiedShufflesplit function in the SciKit-Learn package, which is in the CAD positive dataset and the CAD negative datasethttp://scikit-learn.org/stable/modules/generated/ sklearn.model_selection.StratifiedShuffleSplit.htmlIs described herein, which is incorporated by reference in its entirety. The remaining data sets in the development set that are not used in the validation set are used for training (814 total).

At step 918, the system randomly generates a set of hyper-parameters for the CNN architecture from the search space provided in table 1. These hyper-parameters are specific to the experimental work and were found to cover the interesting range of all parameters in the studied experiments. The system uses a Keras open source neural network library (example code shown in fig. 6) to build a CNN model from a hyper-parametric search space.

At step 920, the system trains the CNN model over multiple time periods (epochs), where each time period includes a single traversal of the entire training set. At the end of each session, the system in the experiment calculated the training AUC and the confirmation AUC. The system uses the observed worst case, rather than the best case, to calculate a score for the time period as the AUC minimum for the current version of the model for training and validation signals for selection of the CNN model.

At step 922, the system terminates training after 10 periods in which no new high scores are observed.

At each run, the system saves the best scoring model along with the corresponding hyper-parameters and corresponding predictions for the validation dataset in step 924.

Fig. 10 is a diagram illustrating a process for selecting a convolutional neural network model configured to non-invasively assess the presence or absence of coronary artery disease in a human, according to an exemplary embodiment.

The highest scoring models, e.g., those having an AUC of 0.58 or higher for males and females, respectively, in the validation set were selected from the model search in the experiment for testing on the validation set. Those models with AUC ≧ 0.57 on the validation set were selected for further evaluation of their performance on a larger gating set. These AUCs were chosen as thresholds as they were found to be the best values that allowed the generation of the required number of predictive models with different characteristics. All models selected for testing on the strobe set are tested simultaneously to avoid bias in the model selection process. In experiments on the performance of the validation set and the gating set, Bootstrap Confidence Intervals (CI) were also calculated using the Matlab R2016b function bootci.

Training label of B2B CNN. The system used training labels derived by Gensini-based scores in experiments (which assigned a score to the dataset that reflects the total burden on the myocardium due to coronary lesions in subjects located and quantified by coronary angiography).

This score defined by Gensini includes a severity weight and a location weight. Coronary lesions are assigned values of 1, 2, 4, 8, 16 and 32 (exponential ratio) according to the respective diameter reductions of 25%, 50%, 75%, 90%, 99% and 100%, according to the severity weight. Coronary lesions are assigned a value between 0.5 and 5 according to the position weight, which reflects the relative impact on the entire myocardium according to their position. For example, if a lesion is located upstream of the coronary circulation (e.g., near the aorta), then the lesion has a greater effect on the circulation in the region of the myocardium than a lesion located downstream (distal) of the circulation. The position with the greatest impact on the loop is assigned a value of 5 and the position with the least impact on the loop is assigned a value of 0.5.

Two Gensini-based scores were evaluated. The first Gensini-based score uses the sum of all weighted scores as a training label for the given data. The second, Gensini-based score, uses only worst-case lesions, i.e., lesions whose severity multiplied by location is the maximum. It was observed that the second modified Gensini score is easier to process for machine learning models. In addition, the system applies a logarithmic operation to the modified Gensini score to change the exponential distribution to a linear distribution, thereby making the tag easier to process by the machine learning model.

B2B Results of CNN experimentsExperimental results of CNN performance are shown in tables 3 and 4 (assessed using gated data sets), table 5 (assessed using validated data sets), and table 6 (assessed using a combination of validated and gated data sets).

Tables 3 and 4 show the performance scores of two models assessed from gated datasets for N213 subjects using an 85% threshold (92 of which, 14 of the 92 were diagnosed with CAD; 121 of which, 55 of the 121 were diagnosed with CAD). Bootstrap confidence intervals ("Cl") are shown in parentheses. The threshold was determined according to the required performance on the validation set, i.e.specificity ≧ 0.65, sensitivity as high as possible. Two models (referred to as "Model 85" and "Model 129") were observed to meet the selection criteria.

TABLE 3

TABLE 4

Table 5 shows the performance scores of two models rated according to the verification set of N-130 using a noise threshold of 85% (58 of which were women, 12 of which were diagnosed with CAD, 72 of which were males, and 35 of which were diagnosed with CAD). The threshold is determined according to the desired performance of the set, i.e.specificity ≧ 0.65, sensitivity as high as possible. Bootstrap CI is shown in parenthesis. As shown, the AUC scores for all these models ranged from 0.62 and 0.65.

TABLE 5

Table 6 shows the performance scores of two models assessed from the validated and gated combined dataset with N343 using a noise threshold of 85% (150 of which were women, of which 150 were diagnosed with CAD; 193 of which were males, of which 193 were 95 with CAD). The threshold is determined according to the desired performance of the set, i.e.specificity ≧ 0.65, sensitivity as high as possible. Bootstrap CI is shown in parenthesis. As shown in table 4, the AUC scores for both models ranged from 0.62 and 0.65.

TABLE 6

B2B CNN modelAs described above, it was observed that the two CNN models (referred to as "Model 85" and "Model 129") met the selection criteria on the validation set. Table 7 shows the hyper-parameters of the two CNN models.

TABLE 7

Positioning convolutional neural networkExperiments performed with data collected from the CADLAD study showed that the exemplary system can detect the location of significant Coronary Artery Disease (CAD) in a subject's specific coronary artery by a Convolutional Neural Network (CNN) trained using a broadband phase gradient biopotential signal dataset. Similar to the B2B CNN model, the wideband phase gradient voltage data is only preprocessed to remove baseline drift, normalize the data ranges, and isolate the acquired data on a beat-to-beat basis for beat analysis. The experiment was performed on three coronary arteries, namely the Left Anterior Descending (LAD), Left Circumflex (LCX) and Right Coronary Artery (RCA).

Method for generating a positioning CNNTo generate a convolutional neural network for experimentation, a training system was developed and used to assess a large number of potential architectures and hyper-parameters through random searches. FIG. 9 shows the processing of the complete model generation pipeline.

As described above, and as shown in fig. 9, the system (e.g., as described with reference to embodiments 110, 110a, 110b) retrieves the raw collected phase signals of the patient from the acquisition measurement device or repository. The system removes unwanted baseline signals from the acquired raw signals to generate a centralized data set. The system uses a second-order forward-reverse filter configured to not introduce any phase distortion (i.e., no phase response). The filter is configured with an effective high-pass cut-off frequency of 0.8 Hz. Separately, the system assesses the signal quality of the acquired signal and excludes from subsequent analysis how the acquired signal does not pass the test.

As described above, the system then down-samples the centered data set from an 8KHz sampling rate to 1KHz using an averaging operator to generate a down-sampled centered data set.

As described above, after the downsampling operation, the system extracts a set of heartbeat segment data sets from the downsampled centralized data set, each heartbeat segment data set including a single isolated complete cardiac cycle. The system uses the Pan-Tompkins algorithm to detect peaks and isolate each complete cardiac cycle for each acquisition channel. The output is a fixed window data set of about 0.75 seconds, centered at the highest amplitude point, encompassing the complete cardiac cycle to provide alignment between all of the heartbeat segment data sets. Sets of cardiac cycles are extracted from each acquired channel using the same process. During the course of the experiment, only the data collected from two of the measurement channels (i.e., the data from channels ORTH1 and ORTH3) were used in the analysis, since the observed period variability noise was observed in one of the three collection channels.

The system then normalizes the numerical range for each extracted heartbeat, as described above. The system normalizes each heartbeat segment data for each channel by dividing the data set by the determined maximum absolute value of the data for a given window, thereby confining the data to a range between-1 and + 1. The system then scales down to +0.5 to-0.5 and adds an offset of 0.5 to adjust the range to 0.0 to 1.0. As a result, the average value of each heartbeat segment data set for each channel is 0.5, ranging between 0.0 and 1.0. Through the normalization and alignment operations, the input of a given CNN receives similar ranges and means of all training data sets, resulting in a stronger classifier. Normalization also makes the signal unitless. The data set collected from the CADLAD study was divided into a development pool and a validation pool.

The system used a data set acquired from 730 patients for training and validation using a first generation phase space recorder (versions 1.0 and 1.1) configured with unipolar broadband phase gradient voltage capture for training and validation; and a data set acquired from 334 patients using a second generation phase space recorder (version 1.2) configured with bipolar broadband phase gradient voltage capture. It is observed that using data sets from different acquisition systems improves the performance of the predictor compared to using data sets from a single hardware type. The system selects the assessed CNN model with AUC ≧ 0.57. The system also used a second validation dataset comprising data collected from 164 patients using a second generation facies space recorder (version 1.2). The system also used a third gated data set comprising data acquired from 243 patients using a second generation phase space recorder (version 1.2).

The system generates a set of hyper-parameters for the CNN architecture from the search space, as provided in table 8. The system uses a modified version of the Keras code shown in fig. 6 to build the CNN model from the hyper-parametric search space. The system trains the CNN model over several time periods, where each time period includes a single traversal of the entire training set. At the end of each session, the system calculates the training and validation AUC. The system uses the observed worst case rather than the best case to calculate a score for the time period as the AUC minimum for the current version of the model for training and validation signals for selection of the CNN model. The system terminates operation after 10 periods in which no new high score is observed. At each run, the system saves the best scoring model along with the corresponding hyper-parameters and corresponding predictions for the validation dataset.

TABLE 8

Hyper-parameter	Candidate value
		Size of batch	1024
Learning rate	0.000272669120574
		Number of convolution layers	4
First convolution layer filter size	17
		Number of filters in first convolution layer	14
Stride length	1
		Number of additional dense layers	1
Size of additional dense layer	10
		Activating a function	‘tanh’
Target	[‘LAD’、‘LCX’、‘RCA’]
		Input frequency	1000Hz
Size of maximum pooling	1
		Discard the	0.224455889694
Loss function	‘mean_squared_error’
		Optimizer	‘Adam’
Input channel	[ORTH1、ORTH2]

Because at least one positive prediction from the CAD model is required to make a positive prediction for the localization model, the system is configured to trigger a prediction from the localization model when a positive prediction is determined from the CAD model.

Training labels for locating CNNOnce the system determines the candidate hyper-parameter set, the system trains the neural network on the training set to learn the binary labels for LAD, LCX, and RCA (e.g., "0" for no disease and "1" for disease). The signature was obtained and evaluated from the patient's angiographic report according to the CADLAD study protocol. For this study, the target was a length-3 vector with binary values for LAD, LCX and RCA. For example, target [1,0,1 ]]The LAD label is 1, the LCX label is 0, and the RCA label is 1. Thus, the prediction of the model is also in the form of a length-3 vector with predictions for LAD, LCX and RCA, respectively.

Results of localized CNN experimentsTable 9 shows the experimental results of localized CNN performance assessed using the gated and validated data sets. Table 9 shows the localization model results for a test set of N411 subjects, with 101 diagnosed with CAD in LAD, 66 diagnosed with CAD in LCX, and 72 diagnosed with CAD in RCA. The result provides statistics of both the overall situation of all three artery predictions combined and the statistics of each individual artery. The system calculates the threshold such that for each case, the sensitivity and specificity (when all three predictions are combined)When in one panel) maximized at values of 75% sensitivity and 65% specificity, resulting in thresholds of about-0.01169, about-0.0311, and about 0.0178 for LAD, LCX, and RCA predictions, respectively. For cases where the CAD prediction was positive, this method produced positive predictions for almost all arteries.

TABLE 9

Model (model)	AUC	Sensitivity of the probe	Specificity of
				Integral body	0.67(0.64，0.70)	0.74(0.68，0.79)	0.60(0.57，0.63)
LAD	0.69(0.64，0.75)	0.76(0.66，0.84)	0.63(0.57，0.69)
				LCX	0.67(0.61，0.72)	0.76(0.63，0.85)	0.58(0.53，0.64)
RCA	0.67(0.61，0.73)	0.72(0.61，0.82)	0.61(0.56，0.66)

To avoid or minimize the likelihood of over-prediction of arterial positivity, the system is configured to select thresholds with values of 0.1116, 0.1596, and 0.1840, respectively, to provide CAD positivity for 72% of LAD, 45% of LCX, and 53% of RCA. In other words, for LAD, the threshold results in a positive prediction as follows: 72% of the subjects in the example were CAD positive in LCAD; 45% of the subjects in the example were CAD positive in LCX; 53% of the subjects in the examples were CAD positive in RCA.

Table 10 shows the results for the whole and single artery for all patients in the test set (N411 patients, 101 of which were diagnosed with CAD in LAD, 66 with CAD in LCX, and 72 with CAD in RCA).

Watch 10

Discussion of locating CNNLocalized CNN studies indicate that women (at least those observed in the CADLAD study) tend to have CAD in a single artery. Furthermore, women (at least those observed in CADLAD studies) often develop CAD in the arterioles. These observations may indicate that it is more difficult to find CAD for women; therefore, even during angiography, women are not adequately diagnosed. These results further indicate that for correct diagnosis of female CAD, a larger, more diverse data set with a higher proportion of diseased women should be used.

Additional experimental data and methods (including visual feature analysis) as performed in the CADLAD study, as well as additional details of the methods described herein, are provided in U.S. provisional application No. 62/907,141, which is incorporated herein by reference in its entirety.

Further, an example integration of B2B CNN and/or localized CNN as described herein to generate predictive scores for indicating the presence of disease, including coronary artery disease, is provided in U.S. provisional application No. 62/907,141. In addition, B2B CNN and/or localized CNN described herein can be used alone or in combination with other methods to characterize pathologies such as LHF, abnormal LVEDP, and the like.

Discussion of the related Art

Neural network models (e.g., deep neural network models, convolutional neural network models as described herein) have predictive capabilities across test sets (i.e., validation sets, and gating sets), and may be used in conjunction with other predictive algorithms to further improve the performance of convolutional neural network models. The convolutional neural network model search method as described herein may produce an algorithm with an AUC of 0.65 or greater. A larger validation set may provide a better measure of the true performance of the model in a larger population. For example, a larger data set may provide more examples of the distribution of each disease, namely LAD only, LCX only, RCA only, LAD/LCX, LAD/RCA, LCX/RCA, and LAD/LCX/RCA. These categories may have different signs of illness — thus a larger data set may provide more training examples for more rigorous studies of each category.

Example computing Environment

FIG. 11 illustrates an exemplary computing environment wherein the illustrative embodiments and aspects may be implemented.

The computing device environment is only one example of a suitable computing environment and is not intended to suggest any limitation as to the scope of use or functionality.

Many other general purpose or special purpose computing device environments or configurations may be used. Examples of well known computing devices, environments, and/or configurations that may be suitable for use include, but are not limited to: personal computers, server computers, hand-held or laptop devices, multiprocessor systems, microprocessor-based systems, network Personal Computers (PCs), minicomputers, mainframe computers, embedded systems, distributed computing environments that include any of the above systems or devices.

Computer-executable instructions, such as program modules, executed by a computer may be used. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. Distributed computing environments may be used where tasks are performed by remote processing devices that are linked through a communications network or other data transmission medium. In a distributed computing environment, program modules and other data may be located in both local and remote computer storage media including memory storage devices.

With reference to FIG. 11, an example system for implementing aspects described herein includes a computing device, such as computing device 1100. In its most basic configuration, computing device 1100 typically includes at least one processing unit 1102 and memory 1104. Depending on the particular configuration and type of computing device, memory 1104 may be volatile (such as Random Access Memory (RAM)), non-volatile (such as read-only memory (ROM), flash memory, etc.), or some combination of the two. This most basic configuration is illustrated in fig. 11 by dashed line 1106.

Computing device 1100 may have additional features/functionality. For example, computing device 1100 may include additional storage (removable and/or non-removable) including, but not limited to, magnetic or optical disks or tape. Such additional storage is illustrated in FIG. 11 as removable storage 1108 and non-removable storage 1110.

Computing device 1100 typically includes a variety of computer-readable media. Computer readable media can be any available media that can be accessed by device 1100 and includes both volatile and nonvolatile media, removable and non-removable media.

Computer storage media includes volatile and nonvolatile, and removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Memory 1104, removable storage 1108 and non-removable storage 1110 are all examples of computer storage media. Computer storage media include, but are not limited to: RAM, ROM, electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, CD-ROM, Digital Versatile Disks (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can be accessed by computing device 1100. Any such computer storage media may be part of computing device 1100.

Computing device 1100 may contain communication connections 1112 that allow the device to communicate with other devices. Computing device 1100 may also have input device(s) 1114 such as keyboard, mouse, pen, voice input device, touch input device, etc. (alone or in combination). Output devices 1116 such as a display, speakers, printer, vibrating mechanism, etc. may also be included, either alone or in combination. All of these devices are well known in the art and need not be discussed at length here.

It should be appreciated that the various techniques described herein may be associated with hardware components or software components, or implemented by a combination of both, where appropriate. Example types of hardware components that may be used include Graphics Processing Units (GPUs), Field Programmable Gate Arrays (FPGAs), Application Specific Integrated Circuits (ASICs), Application Specific Standard Products (ASSPs), systems on a chip (SOCs), Complex Programmable Logic Devices (CPLDs), and the like. The methods and apparatus of the presently disclosed subject matter, or certain aspects or portions thereof, may take the form of program code (i.e., instructions) embodied in tangible media, such as floppy diskettes, CD-ROMs, hard drives, or any other machine-readable storage medium, wherein, when the program code is loaded into and executed by a machine, such as a computer, the machine becomes an apparatus for practicing the presently disclosed subject matter.

Although an illustrative implementation may refer to the use of the presently disclosed subject matter in the context of one or more stand-alone computer systems, the subject matter is not so limited, but may be implemented in any computing environment, such as a network or distributed computing environment. Moreover, various aspects of the presently disclosed subject matter may be implemented in or across a plurality of processing chips or devices, and the storage may similarly be implemented across a plurality of devices. For example, these devices may include personal computers, network servers, handheld devices, and wearable devices.

Although the subject matter has been described in language specific to structural features and/or methodological acts, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts described above are disclosed as example forms of implementing the claims.

Other examples of processes that may be used with the illustrative methods and systems are described in the following documents: U.S. Pat. No. 9,289,150 entitled "Non-innovative Method and System for manipulating Cardiovasular Systems"; U.S. Pat. No. 9,655,536 entitled "Non-innovative Method and System for manipulating Cardiovasular Systems"; U.S. Pat. No. 9,968,275 entitled "Non-innovative Method and System for manipulating Cardiovasular Systems"; U.S. Pat. No. 8,923,958 entitled "System and Method for Evaluating an electrophoretic Signal"; U.S. Pat. No. 9,408,543 entitled "Non-innovative Method and System for manipulating Cardiovasular Systems and All-Cable monitor and Sudden Cardioc Death Risk"; U.S. Pat. No. 9,955,883 entitled "Non-innovative Method and System for manipulating Cardiovasular Systems and All-Cable monitor and Sudden Cardioc Death Risk"; U.S. Pat. No. 9,737,229 entitled "Noninrivative Electropathological Method for Estimating Mammarian Cardiac Chamber Size and Mechanical Function"; U.S. Pat. No. 10,039,468 entitled "Noninrivative Electropathological Method for Estimating Mammarian Cardiac Chamber Size and Mechanical Function"; U.S. Pat. No. 9,597,021 entitled "Noninrivative Method for Estimating Glucose, Glycosylated Hemoglobin and Other Blood compositions"; U.S. Pat. No. 9,968,265 entitled "Method and System for manipulating Cardiovasular Systems From Single Channel Data"; U.S. Pat. No. 9,910,964 entitled "Methods and Systems Using chemical Analysis and Machine Learning to diagnosis" for the purposes of the present invention; U.S. patent application publication No. 2017/0119272 entitled "Method and Apparatus for Wide-Band Phase Gradient Signal Acquisition"; PCT publication No. WO2017/033164, International application for Wide-Band Phase Gradient Signal Acquisition, entitled "Method and Apparatus"; U.S. patent application publication No. 2018/0000371 entitled "Non-innovative Method and System for Measuring Myocardial Ischemia, Stenosis Identification, Localization and Fractional Flow Reserve Estimation"; PCT publication No. WO2017/221221, International application entitled "Non-innovative Method and System for Measuring Myocardial Ischemia, Stenosis Identification, Localization and Fractional Flow Reserve Estimation"; U.S. Pat. No. 10,292,596 entitled "Method and System for Visualization of Heart Tissue at Risk"; U.S. patent application No. 16/402,616 entitled "Method and System for Visualization of Heart Tissue at Risk"; U.S. patent application publication No. 2018/0249960 entitled "Method and System for Wireless Phase Gradient Signal Acquisition"; U.S. patent application No. 16/232,801 entitled "Method and System to Assembly Disease Using Phase Space volume Objects"; international application No. PCT/IB/2018/060708 entitled "Method and System to Assembly Disease Using Phase Space volume Objects"; U.S. patent application publication No. US2019/0117164 entitled "Methods and Systems of De Noisying Magnetic-Field Based Sensor Data of Electrical Signals"; a U.S. patent application publication No. 2019/0214137 entitled "Method and System to Assembly Disease Using Phase Space Tomography and Machine Learning" filed on 26.12.2018; PCT International application No. PCT/IB2018/060709, entitled "Method and System to Assembly Disease Using Phase Space Tomography and Machine Learning"; U.S. patent application publication No. 2019/0384757 entitled "Methods and Systems to quantitative and removable anaerobic Noise in Biophysical Signals" filed on 18.6.2019; U.S. patent application Ser. No./___________, entitled "Method and System to Assembly Disease Using Phase Space Tomography and Machine Learning", filed concurrently herewith (attorney docket No. 10321 and 034usl, and claiming priority from U.S. patent provisional applications Nos. 62/784,984 and 62/835,869); U.S. patent application publication No. US2019/0365265 entitled "Methods and Systems to Assembly Phase Space dynamics and Machine Learning"; U.S. patent application Ser. No./______, entitled "Method and System for Automated Quantification of Signal Quality" (attorney docket No. 10321- > 036usl and claiming priority from U.S. provisional patent application No. 62/784,962); U.S. patent application No. 15/653,433 entitled "discovery Novel Features to Use in Machine Learning Techniques, Such as Machine Learning Techniques for marking Medical Condition"; U.S. patent application No. 15/653,431 entitled "discovery genes to Use in Machine Learning Techniques"; U.S. patent application Ser. No./__ (attorney docket No. 10321-040pvl claims priority to U.S. provisional patent application No. 62/862,991), entitled "Method and System to Association diseases Using Dynamic Analysis of biological Signals"; application Ser. No./__ (attorney docket No. 10321- _ 041pvl and claiming priority from U.S. provisional patent application No. 62/863,005), U.S. provisional patent application entitled "Method and System to Assembly diseases Using dynamic Analysis of Cardiac and Photophythographic Signals", each of which is incorporated herein by reference in its entirety.

Unless explicitly stated otherwise, it is not intended that any method set forth herein be construed as requiring that its steps be performed in a particular order. Thus, if a method claim does not actually recite an order to be followed by its steps or it is not otherwise specifically stated in the claims or descriptions that the steps are to be limited to a specific order, it is no way intended that such order be inferred, in any respect. This applies to any possible non-explicit basis for interpretation, including: logic issues regarding step arrangements or operational flows; simple meaning derived from grammatical organization or punctuation; the number or type of embodiments described in the specification.

While the methods and systems have been described in connection with certain embodiments and specific examples, it is not intended to limit the scope to the particular embodiments set forth, as the embodiments herein are intended in all respects to be illustrative rather than restrictive.

The methods, systems, and processes described herein can be used to generate stenosis and FFR outputs for use with procedures such as placing a stent-graft within a vessel, such as an artery of a living (e.g., human) subject, as well as other interventional and surgical systems or processes. In one embodiment, the methods, systems, and processes described herein may be configured to use FFR/stenosis output to determine and/or modify the number of stents to be placed in a living body (e.g., a human) intra-operatively, including their optimal deployment location within a given vessel, and the like.

Examples of other biophysical signals that may be analyzed in whole or in part using the exemplary methods and systems include, but are not limited to: an Electrocardiogram (ECG) dataset, an electroencephalography (EEG) dataset, a gamma-synchronized signal dataset, a respiratory function signal dataset, a pulse oximetry signal dataset, a perfusion data signal dataset, a quasiperiodic biosignal dataset, a fetal ECG dataset, a blood pressure signal, a cardiac magnetic field dataset, and a heart rate signal dataset.

Exemplary analyses are useful for diagnosis and treatment of heart-related pathologies and disorders and/or nervous system-related pathologies and disorders, and such evaluations are applicable to diagnosis and treatment (including surgery, minimally invasive and/or drug therapy) of any pathology or disorder in which biophysical signals are related to any relevant system of a living body. One example of a cardiac aspect is the diagnosis of CAD and its treatment (alone or in combination) by various therapies, such as placement of stents in coronary arteries, performing atherectomy, angioplasty, prescription of medications, and/or prescription of exercise, nutrition, and other lifestyle changes. Other heart-related pathologies or conditions that may be diagnosed include, for example, cardiac arrhythmias, congestive heart failure, valve failure, pulmonary hypertension (e.g., pulmonary hypertension due to left heart disease, pulmonary hypertension due to pulmonary disease, pulmonary hypertension due to chronic thrombosis, and pulmonary hypertension due to other diseases (e.g., blood or other diseases)), and other heart-related pathologies, conditions, and/or diseases. Non-limiting examples of diagnosable nerve-related diseases, pathologies or conditions include, for example, epilepsy, schizophrenia, Parkinson's disease, Alzheimer's disease (and all other forms of dementia), autism spectrum (including Alzheimer's syndrome), attention deficit hyperactivity disorder, Huntington's disease, muscular dystrophy, depression, bipolar disorder, brain/spinal cord tumors (malignant and benign), movement disorders, cognitive disorders, language disorders, various psychoses, brain/spinal cord/nerve injury, chronic traumatic encephalopathy, cluster headache, migraine, neuropathy (various forms including peripheral neuropathy), phantom limb/pain, chronic fatigue syndrome, acute and/or chronic pain (including back pain, back surgery failure syndrome, etc.), motor dysfunction, anxiety disorders, diseases caused by infection or extrinsic factors (such as Lyme disease, Huntington's disease, muscular dystrophy), Encephalitis, rabies), narcolepsy and other sleep disorders, post-traumatic stress disorder, neurological disorders/effects associated with stroke, aneurysm, hemorrhagic damage, and the like, tinnitus and other hearing related diseases/disorders, and vision related diseases/disorders.

When any value or range is described herein, unless clearly stated otherwise, that value or range is approximate. When any range is described herein, unless clearly stated otherwise, that range includes all values therein and subranges therein. Any information in any material (U.S. patent/foreign patent, U.S. patent application/foreign patent application, book, article, etc.) that is incorporated by reference herein is incorporated by reference only to the extent that there is no conflict between such information and the other statements and drawings set forth herein. In the event of a conflict, including one that would render any claim herein invalid or that would not require priority, then any such conflicting information incorporated by reference material is specifically not incorporated by reference.