Method for diagnosing respiratory system disease by analyzing cough sound using disease characteristics
阅读说明:本技术 用疾病特征分析咳嗽声音以诊断呼吸系统疾病的方法 (Method for diagnosing respiratory system disease by analyzing cough sound using disease characteristics ) 是由 U·阿贝拉特纳 V·斯旺加 于 2018-12-20 设计创作,主要内容包括:一种诊断患者一种或多种呼吸道疾病的方法,包括以下步骤:从患者获得咳嗽声音;对所述咳嗽声音进行处理,以产生表示来自咳嗽片段的一个或多个咳嗽声音特征的咳嗽声音特征信号;基于咳嗽声音特征信号获得一个或多个疾病特征;以及将一个或多个疾病特征分类,以将咳嗽片段视为指示一种或多种所述疾病;其中,基于咳嗽声音特征信号获取一个或多个疾病特征的步骤包括:将咳嗽声音特征应用于一个或多个预先训练的疾病特征决策机中的每一个,每一个所述决策机已预先训练以将咳嗽声音特征分类为对应于特定疾病或非疾病状态,或对应于第一特定疾病或与第一特定疾病不同的第二特定疾病。(A method of diagnosing one or more respiratory diseases in a patient comprising the steps of: obtaining a cough sound from a patient; processing the cough sound to produce a cough sound signature signal representing one or more cough sound signatures from a cough segment; obtaining one or more disease characteristics based on the cough sound characteristic signal; and categorizing one or more disease features to treat a cough segment as indicative of one or more of said diseases; wherein the step of obtaining one or more disease signatures based on the cough sound signature signal comprises: applying the cough sound signature to each of one or more pre-trained disease signature decision machines, each of the decision machines having been pre-trained to classify the cough sound signature as corresponding to a particular disease or non-disease state, or to a first particular disease or a second particular disease different from the first particular disease.)
1. A method of diagnosing one or more respiratory diseases in a patient comprising the steps of:
obtaining a cough sound from a patient;
processing the cough sound to produce a cough sound signature signal representing one or more cough sound signatures from a cough segment;
obtaining one or more disease characteristics based on the cough sound characteristic signal; and
classifying one or more disease features to treat a cough segment as indicative of one or more of said diseases;
wherein the step of deriving one or more disease characteristics based on the cough sound characteristic signal comprises applying the cough sound characteristics to each of one or more pre-trained disease characteristic decision machines, each of which has been pre-trained to classify the cough sound characteristics as corresponding to a disease state or a non-disease state, or corresponding to a first specific disease or a second specific disease different from the first specific disease.
2. The method of claim 1, wherein the one or more disease feature decision machines each comprise a trained Logistic Regression Model (LRM).
3. The method of claim 2, wherein each trained Logistic Regression Model (LRM) is trained using a reduced training feature set, the features determined to be important to the LRM, thereby avoiding over-training of the LRM.
4. The method of claim 2A, wherein the P-value is selected by calculating an average P-value for all training features and then selecting the average P-value to be less than a threshold PminDetermining training features is important to the LRM.
5. The method of any of claims 2 to 4, wherein the argument for each LRM is a value of a cough sound signature, and the output value of the LRM includes a predicted probability of cough indicative of a first particular disease referenced to a second particular disease or to a non-disease state.
6. The method of claim 1, wherein the one or more disease feature decision machines comprise one or more trained neural networks.
7. The method according to any one of the preceding claims, comprising applying clinical patient measurements as arguments to a disease signature decision machine in addition to the cough signature.
8. The method of any one of the preceding claims, wherein the step of categorizing one or more disease features to treat a cough segment as indicative of one or more of the diseases comprises: applying the one or more disease features to a single classifier trained to encompass all disease groups.
9. The method of any one of claims 1 to 7, wherein the step of categorizing one or more disease features to treat a cough segment as indicative of one or more of the diseases comprises: applying the one or more disease features to a plurality of classifiers, each of the plurality of classifiers trained to identify a disease of interest.
10. The method of claim 86 or 9, wherein the step of classifying the one or more disease features comprises classifying the one or more disease features as one of the following diseases: bronchiolitis (S)Bo) (ii) a Laryngitis (S)C) (ii) a asthma/RAD (S)A) (ii) a Pneumonia (S)P) (ii) a Lower respiratory tract disease (S)LRTD) (ii) a Primary upper respiratory tract infection (S)U)。
11. The method of claim 9, wherein d disease features are generated by the step of obtaining one or more disease features based on the cough sound signature signal, and accordingly the Artificial Neural Network (ANN) has a d-dimensional input layer with one input neuron corresponding to each disease feature.
12. The method of claim 11, wherein the artificial neural network has a k-dimensional output layer, wherein each neuron in the output layer outputs a probability corresponding to a disease.
13. The method of any one of the preceding claims, wherein the step of classifying one or more disease features comprises combining a composite effective probability metric PQ' compile to:
P'Q=PQ(1-PZ),
wherein P isQIndicator comprising the probability of a patient belonging to disease Q, PZIndicator comprising the probability of a patient belonging to a disease Z, wherein the product PQ(1-PZ) Representing the probability of a compound event that the patient belongs to disease Q but not to disease Z.
14. The method of claim 10, comprising calculating a cough index for each target disease, wherein the cough index for a patient target disease is calculated as a ratio of the number of coughs for a patient classified as indicative of a target disease to the total number of coughs analyzed for the patient.
15. The method of any preceding claim, comprising applying one or more post-screening tests to the classification from the step of classifying one or more disease features to detect a dominant false positive.
16. The method of claim 15, further comprising the step of adjusting the classification of one or more disease features based on the detected false positives.
17. The method of any of the above claims, further comprising applying a specific treatment to the patient based on the specific disease diagnosed.
Technical Field
The present invention relates to methods and devices for assisting medical personnel in diagnosing and treating patients suffering from respiratory diseases.
Background
The reference to a prior art method, apparatus or document is not an admission that it forms or forms part of the common general knowledge.
In one or more of the same inventors' previous work, which was the subject of international patent application PCT/AU2013/000323 (which is herein incorporated by reference in its entirety), respiratory tract sounds of patients were recorded and cough sounds were identified therefrom. Features are extracted from the cough sounds to form test feature vectors, which are then applied to a pre-trained classifier, preferably a logistic regression model, to diagnose the presence of respiratory dysfunction, such as pneumonia, in the patient.
Although the method of diagnosing a disease state described in PCT/AU2013/000323 works well and has been successfully commercially implemented, there is still a need for improvement. For example, it would be advantageous if a method could be provided that improved in that more accurate diagnostic results could be produced. It is also preferred that the method can inject domain specific information into the diagnostic process. Furthermore, it would be desirable if the subjective nature of standard clinical diagnosis could be accommodated.
The invention aims to provide a method and a device for assisting in diagnosing respiratory disease states.
Disclosure of Invention
According to a first aspect of the present invention there is provided a method of diagnosing one or more respiratory diseases in a patient comprising the steps of:
obtaining a cough sound from a patient;
processing the cough sound to produce a cough sound signature signal representing one or more cough sound signatures from a cough segment;
obtaining one or more disease signatures based on the cough sound signature signal; and
classifying one or more disease features to treat a cough segment as indicative of one or more of said diseases;
wherein the step of deriving one or more disease characteristics based on the cough sound characteristic signal comprises applying the cough sound characteristics to each of one or more pre-trained disease characteristic decision machines, each of which is pre-trained to classify the cough sound characteristics as corresponding to a particular disease or non-disease state, or to a first particular disease or a second particular disease different from the first particular disease.
In a preferred embodiment of the invention, the one or more disease feature machines each comprise a trained Logistic Regression Model (LRM).
Each trained Logistic Regression Model (LRM) may be trained using a reduced set of training features, the features of the available training feature set being determined to be important for the LRM, thereby avoiding over-training of the LRM.
By calculating the average P-value of all training features and then selecting the average P-value to be less than the threshold value PminDetermining that the training features are significant for the LRM. Further details of this process according to one embodiment of the present invention are set forth in appendix F herein.
In this embodiment, the argument of each LRM is a value of a cough sound characteristic, and the output value of the LRM includes a predicted probability of cough, which indicates a first specific disease referring to a second specific disease or referring to a non-disease state.
The disease features may also be trained to produce scores that correlate to particular assessments or measurements used in respiratory medicine. Example (c): wheeze Severity Score (WSS) that distinguishes low WSS from high WSS by training features; low FEV1 and high FEV1 as measured in spirometry; low FEV1/FVC and high FEV1/FVC measured in spirometry.
In another embodiment of the present invention, the one or more disease signature machines may include one or more trained neural networks. Other classifiers or models that provide a continuous output (e.g., generalized linear models, hidden markov models, etc.) may also be used as semaphores.
The method may further comprise applying the clinical patient measurement as an argument to a disease characteristic decision machine in addition to the cough characteristic.
In one embodiment of the invention, the step of categorizing one or more disease features to treat a cough segment as indicative of one or more of said diseases comprises: one or more disease features are applied to a single classifier that is trained to cover all disease groups.
In another embodiment of the present invention, the step of classifying one or more disease features to treat a cough segment as indicative of one or more of said diseases comprises: one or more disease features are applied to a plurality of classifiers, each classifier being trained to identify a disease of interest.
In a preferred embodiment, each classifier is trained to identify one of the following diseases: bronchiolitis (S)Bo) (ii) a Laryngitis (Croup) (S)C) (ii) a asthma/RAD (S)A) (ii) a Pneumonia (S)P) (ii) a Lower respiratory tract disease (S)LRTD) (ii) a Primary upper respiratory tract infection (S)U)。
Preferably, the d disease features are generated by the step of obtaining one or more disease features based on the cough sound feature signal, and accordingly the Artificial Neural Network (ANN) has a d-dimensional input layer with one input neuron corresponding to each disease feature.
In a preferred embodiment of the invention, the ANN has a k-dimensional output layer, wherein each neuron in the output layer outputs a probability corresponding to a disease.
In a preferred embodiment of the invention, the step of classifying one or more disease features comprises compiling a composite significance probability metric y compiling a composite significance probability metric P'QComprises the following steps:
P′Q=PQ(1-PZ)
wherein P isQIncluding an indicator of the probability of a patient belonging to disease Q, PZIncluding an indicator of the probability of a patient belonging to disease Z. Thus, the product PQ(1-PZ) Representing the probability of a compound event that the patient belongs to disease Q but not to disease Z.
The method may include calculating a cough index for each target disease, wherein the cough index for a patient of a target disease is calculated as a ratio of "cough of a patient classified as indicative of a target disease" to "total number of coughs analyzed for said patient".
The method may also include applying a specific therapy, such as a therapy known to be effective for the patient based on the specific disease being diagnosed.
Drawings
Preferred features, embodiments and variants of the invention can be discerned from the following detailed description, which provides sufficient information for a person skilled in the art to carry out the invention. The detailed description is not to be taken as limiting the scope of the foregoing summary in any way. The detailed description will refer to the following figures:
fig. 1 is a flow chart of a preferred diagnostic method according to a first embodiment of the invention.
FIG. 2 depicts an artificial neural network structure with a d-dimensional input layer (input from feature blocks in diagram B) and a k-dimensional output layer. Each neuron in the output layer corresponds to a subset of diseases.
Fig. 3 is a first portion of a flow diagram of a diagnostic method according to another embodiment of the invention, which includes a pre-screening process block and a post-screening process block.
Fig. 4 includes a second portion of the flow chart of fig. 3.
FIG. 5 is a block diagram of an auto-encoder neural network trained for feature mapping.
FIG. 6 is a block diagram of a feature layer neural network, which is trained to generate disease-specific features.
FIG. 7 is a block diagram of a deep neural network, in accordance with one embodiment of the present invention.
Detailed Description
Fig. 1 depicts a
The
(i) can provide good training and verification performance in the personal feature generation process,
(ii) domain specific knowledge (e.g., if the goal of the final classifier is to diagnose asthma/RAD, WSS features would be useful, since wheezing is a powerful indicator of asthma/RAD; if the classifier has to diagnose pneumonia, it is possible to indicate pneumonia and bronchiolitis features, since these are known to represent diagnostic predicaments in clinical practice, etc.),
(iii) based on a search process on the features to maximize the diagnostic performance of the training/validation set classifier output.
The output values from the selected feature selection block 116 are passed to a
A diagnostic instructions block 120 is responsive to the
The following discussion will provide further details.
Cough
At
Furthermore, in accordance with a preferred embodiment of the present invention, the cough
The cough signature calculation block is arranged to perform the following process in use:
(i) let x denote the discrete-time sound signal from any cough event.
(ii) X is divided, for example, into three equally sized, non-overlapping subfragments. The objective goal is to capture the changes in the mathematical characteristics in one cough. Let x beiRepresents the ith subfragment of x, wherein i ═ 1, 2 and 3.
From each subfragment xiThe following characteristics were calculated: 8 Bispectral Coefficients (BC), non-gaussian fractions (NGS), first four Formant Frequencies (FF), log energy (LogE), Zero Crossings (ZCR), kurtosis (Kurt), 31 mel-frequency cepstral coefficients (MFCC), shannon entropy (SHE).
(iii) In addition, 13 wavelet features were calculated using the entire cough event data [2 ]. For a brief description of each cough signature, see appendix B.
(iv) A total of C was extracted from each cough eventf157 features.
Disease
In previous work (i.e., the topic of PCT/AU 2013/000323), patients were directly classified by training a logistic regression model on the feature sets calculated in
In contrast to the approach taken in PCT/AU2013/000323, one preferred embodiment of the present invention relates to a new process that has been conceived and is referred to herein as "feature generation", the purpose of which is as follows: (a) as a means of injecting domain-specific information into the diagnostic process, (b) as a means of adapting to the subjective nature of standard clinical diagnosis, (c) as a means of producing a more accurate diagnosis.
The feature generation performed at
In one embodiment, the input features (features) of
Logistic Regression (LR) models are generalized linear models that use multiple independent variables to estimate the probability of a classification event. The relevant independent variables are mathematical features calculated from the cough events and the classification events are disease subgroups. Thus, in the above example, the LR model can be trained to predict the probability of cough that belongs to bronchiolitis disease with reference to laryngitis disease. The LR model is derived using a regression function to estimate the probability Y taking into account the independent features, as follows:
z=β0+β1·q1+β2q2+…+βn-1qF(2)
in (2), β0Called intercept, β1、β2Called the regression coefficient, q, representing the independent variable1、q2、…qFThe features are represented.
In order to generate the
Cough sound capture
Cough sounds were recorded from two clinical sites, the Joondalup Health Campus (JHC) and the margarite princess hospital (PMH), both in perose city, western australia. The patient population includes children 0-12 years old and is suspected of having respiratory diseases such as pneumonia, asthma/RAD (reactive airway disease), bronchiolitis, laryngitis and upper respiratory infection (URTI). Human ethics committees of the university of queensland, the Joondalup health campus, and the margarite princess hospital approved the study protocol and patient recruitment procedure.
Patients who met inclusion criteria (with cough, asthma, shortness of breath, wheezing, upper respiratory infection) and did not meet exclusion criteria (required respiratory support, not consented) were enrolled into the study. Healthy subjects, defined as children who did not have any symptoms of respiratory disease at the time of measurement, were also enrolled.
Cough sounds were recorded using apple iPhone 6s (see
Database and experimental design
The database used to train the various models required to implement the apparatus of fig. 1 includes cough records and detailed clinical diagnostic information for each patient, including final diagnosis, clinical examination findings, and laboratory results as well as imaging results. Demographic information in patient de-identified format is also available.
Diagnostic panel(case definitions for diagnosing disease see appendix A).
Normal group (Nr): healthy volunteers with no recognizable respiratory disease were measured.
Primary upper respiratory tract infection group (U): patients with only Upper Respiratory Tract Infections (URTI), no medically identifiable lower respiratory tract involvement or other respiratory tract disease at the time of measurement.
Laryngitis group (C): the diagnosis was classified as patients with laryngitis alone or with upper respiratory infection.
Asthma/reactive airway disease group (a): the diagnosis is classified as patients with asthma or reactive airway disease, with or without upper respiratory tract infections.
Clinical pneumonia group (P): the diagnosis is classified into patients with clinical pneumonia with or without upper respiratory tract infection.
Bronchiolitis group (Bo): patients classified as bronchiolitis with or without upper respiratory infection were diagnosed.
Bronchitis group (Bc): patients diagnosed with bronchitis with or without upper respiratory infection.
All subjects are divided into two mutually exclusive sets, and the classifier model is trained, verified and tested. The two mutually exclusive sets are: (1) training validation set (TrV) and (2) prospective test set (PT). Each subject belongs to only one set. Subjects with diagnostic uncertainty (as shown by the clinical team) and co-morbidities (except for upper respiratory infection) were excluded from TrV.
The training validation set TrV is used to train and validate the model following leave-one-out-validation (LOOV) or K-fold cross-validation techniques. The LOOV method involves training the model using data from all but one patient and validating the model using the cough events of the remaining patients. This process is systematically repeated so that each patient in TrV is used to accurately validate the model once. In K-fold cross-validation, the original sample is randomly divided into K equally sized subsamples. A single subsample is retained as validation data for the test model. The remaining (K-1) samples will be used to train the model. This process will be repeated K times until all the data in TrV is used once when testing the model. Note that LOOV is a special case of a K-fold cross-validation method, with K set to the total number of data (N) in set TrV.
Table 1 (spread) lists the
Table 1: the LOOV embodiment generates a list of models for the features trained in the feature block of FIG. 2.
TrV the data in the dataset was used to train and validate all of the LR models listed in table 1.
In the simplest form, the LR model is trained using only cough-based features. However, the inventors have recognized the existence of some simple clinical measures that can be used to improve the performance of the LR model, with minimal complexity and no additional cost. Inspired by this, the inventors attached a cough-based feature to a simple clinical feature and trained the second set of LR model lists in table 1. Table 2 shows a simple clinical profile of the additional cough-based profile.
TABLE 2: "√" indicates that the feature is included in the model design. Clinical symptoms were extracted from the patient's medical history according to the parent's report at the time of clinical examination.
Feature selection
Feature selection is a technique for selecting relevant features for designing an optimal feature model. For example, when constructing an LR model as a feature using the TrV set, a p-value would be calculated for each input feature to capture the importance of the particular feature to the model. The p-value of the important features is lower. This property of the LR model is applied to the entire TrV set to select reasonable feature combinations. Once the subset of significant features is known, it is used to retrain the LR model based on the LOOV (K-fold) training/validation of the TrV dataset. More details of the feature selection process may be found in earlier PCT applications and "abeyratase, u.r. et al, Cough soundadalysis can rapid diagnosis reagent pulmonary research, annals of biomedicalizing, 2013, 41 (11): 2448 and 2462.
Selection of good LR model
LOOV training/validation Process produces N for LR modelkNumber, in which NkTrV are the numbers of patients in the pool. Because TrV patients were counted differently for different disease groups, N for the different LR models listed in Table 1kThe values will also differ. From NkIn the LR model, an optimal model is selected based on a k-means clustering algorithm. For more details on model selection using k-means clustering algorithm, see [1]]。
Is provided with
Represents selected features of an LR model j trained based on the use of cough-only features, andrepresenting a selected LR model trained using coughing and simple clinical features. Once the LR model is selected, it will be run in all patients in the TrV dataset to generate disease signatures.These characteristics provide the "response" of a given patient i to all hypotheses tested by different characteristic values "
Andvector XiRepresents a modelAndcough of the i patient in need thereofCollecting information such as cough sound and clinical symptoms.For example, if the patient being tested suffers from bronchiolitis, his/her response to the model (e.g. { bronchiolitis with all other diseases }) should give a strong value close to 1 ("complete response"), whereas { A/RAD with all other diseases } should give a lower response ("partial response"). However, the patient will give a specific response for each characteristic axis (LRM model) and a vector V representing the set of these responses for each patient ii,cAnd Vi,cfWill characterize the disease from which the patient is suffering.
Vi,1=[{ρc,ijJ ═ all selected features (LRM model)](3)
Vi,cf=[{ρc,ijJ ═ all selected features (LRM model)](4)
Not all
The function of the
Fig. 2 shows an exemplary structure 200 of the softmax ANN used in
P′Q=PQ(1-PZ) (5)
wherein P isQCan be regarded as an indicator of the probability that the patient belongs to the disease Q, PZCan be considered as an indicator of the probability that the patient belongs to disease Z. Thus, the product PQ(1-PZ) Representing the probability of a compound event that the patient belongs to disease Q and not to disease Z.
The ANN shown in fig. 2 shows an artificial neural network structure having a d-dimensional input layer 201 and a k-dimensional output layer 205, the d-dimensional input layer 201 being composed of neurons 203 (receiving input from the
In a preferred embodiment of the invention, five different softmax-ANN models are trained to identify disease subgroups: bronchiolitis, laryngitis, asthma/RAD, pneumonia, and Lower Respiratory Tract Disease (LRTD). Details of these disease-specific softmax-ANN are provided below.
1. Capillary bronchitis softmax-ANN (S)Bo): in this softmax-ANN model, features centered on bronchiolitis were used. The characteristic LR model for training ANN isIn the output layer, the k-dimension is set to 6 neurons, each neuron corresponding to a disease subgroup, bronchiolitis, asthma/RAD, pneumonia, bronchitis, pURTI, laryngitis. In the formula (5), PQIs the output of bronchiolitis neurons, PZIs the output of the RAD neuron.
2. Softmax (S) of laryngitisC): in thatIn this ANN model, features centered on laryngitis were used. The characteristic LR model used isIn the output layer, the k-dimension is set to 6 neurons, each neuron corresponding to a disease subgroup, bronchiolitis, asthma/RAD, pneumonia, bronchitis, pURTI, laryngitis. In formula (5), PQIs the output of the laryngitis neuron, PZIs the output of the pURTI neuron.
3. softmax-ANN (S) for asthma/RADA): in this ANN model, features centered on asthma/RAD were used. The characteristic LR model used is
In the output layer, the k-dimension is set to 6 neurons, each neuron corresponding to a disease subgroup, bronchiolitis, asthma/RAD, pneumonia, bronchitis, pURTI, laryngitis. In the formula (5), PQIs the output of asthma/RAD neurons, PZIs the output of the pURTI neuron.4. softmax-ANN (S) of pneumoniaP): in the pneumonia softmax-ANN model, the characteristic LR model used is
At the output layer, the k dimension is set to 4 neurons; one neuron for bronchiolitis, pneumonia and laryngitis, respectively, and one neuron for asthma/RAD, bronchitis and pURTI. In formula (5), PQIs the output of the pneumonia neuron, PZIs the output of neuron of asthma/RAD, bronchitis and pURTI disease.5. softmax-ANN (S) for lower respiratory tract diseaseLRTD): lower Respiratory Tract Disease (LRTD) is a covered term used to denote involvement of the call-downDiseases of the aspiration tract. The LRTD subgroup combines patients with pneumonia, asthma/RAD, bronchiolitis, and bronchitis diseases. The LRTD soft-ANN model was trained to identify LRTD disease from laryngitis and pURTI. In this ANN, the LRTD-centered feature is used. The characteristic LR model used is
On the output layer, the k-dimension is set to 3 neurons, 1 neuron each for URTI and houy, and 1 neuron for LRTD disease group. In formula (5), PQIs the output of the LRTD neuron, PZIs the output of the pURTI neuron.6. Softmax ANN (S) of primary URTIU): in this ANN model, pURTI-centered features were used. The characteristic LR model used is
In the output layer, the k-dimension is set to 6 neurons, each neuron corresponding to a disease subgroup, bronchiolitis, asthma/RAD, pneumonia, bronchitis, pURTI, laryngitis. In the formula (5), PQIs the output of the pURTI neuron, PZIs the output of the pneumonia neurons.To train the cough-only softmax ANN, an LR signature model developed with only the cough signature was used. To train softmax ANN for cough plus simple clinical features, LR feature models developed with cough features only and/or cough plus simple clinical features were used.
Selecting a preferred softmax-ANN model
The LOOV cross-validation process will generate N softmax models. Where N is the number of patients in the TrV data set. From the N models, the inventors again selected an optimal model according to the k-means clustering algorithm. For more details on model selection using k-means clustering algorithm, please refer toSee [1]]. Let SsIndicates the selected softmax ANN, λsThe threshold is decided for the corresponding probability for a particular disease subgroup. Once S is selectedsAll parameters of the model are fixed and the training process is completely terminated. Then model SsUsed as the best model for further testing. A similar approach was used in the K-fold cross-validation method.
Other factors that the inventors explore when training softmax ANN are the size of the network, training time, training rate, stopping criteria, and the difference between training and validation errors. The purpose of studying these factors is to minimize the over-learning of the data set that we can utilize and to maximize the spread to statistically similar, previously unseen populations.
Several preferred embodiments of the method for making a diagnostic decision for each patient are described below.
i) Cough index and patient classification
In this method, the directness of the classifier (e.g. P) is processedQ、PZEtc.) output (e.g., softmax ANN) or probabilistic composite metric (e.g., P'QEtc.). The first goal is to test patients for each cough according to the target disease category (e.g., bronchiolitis) and label each cough as "1 successful" or "0 failed". For example, if the ultimate goal is to detect bronchiolitis, each cough is flagged as bronchiolitis (success, flagged as 1) or non-bronchiolitis (failure, flagged as 0). To select the optimal decision threshold λ (i.e., if the test statistic ≧ λ, the cough in the test belongs to the target disease (success, 1)), a Receiver Operating Curve (ROC) analysis was used on set TrV.
The cough index for each target disease was then calculated as shown below. Is provided with CTIs the total number of coughs, C, analyzed from patient iSIs the number of coughs marked as "successful" using the softmax-ANN model. Then, for the target disease subgroup j, the cough index CI for patient ii,jThe calculation is as follows: CIi,j=CS/CT。
ii) deep learning strategy based classification
The raw output of the softmax layer, which is a real number with continuous variation between 0 and 1, can be further processed in the spirit of the deep learning method. Some preferred embodiments are given below:
a) the raw output of the Softmax ANN of block 1183 may be fed to another classifier, such as a neural network, which will simultaneously receive input from other similar networks trained on different principles. For example, another network may be based entirely on clinical symptoms that parents can easily observe and report to clinicians. The clinical symptoms network may have a parallel structure to the disease characteristics generation block 114 of fig. 1, or a simpler version, such as a set of selected LRM classifiers (or ANN classifiers, etc.) that map clinical symptoms to a desired disease group.
The LRM based
In a preferred embodiment, the inventors trained a deep neural network using an encoder method. The final classifier is itself a neural network, where one neuron represents each disease class of interest.
D. Test procedure
In this section, we will discuss the performance of the final diagnostic model on the expected data set. Before using the expected data set, we completely freeze our diagnostic model, allowing no further training, parameter tuning or protocol modification.
Results
A. Data set
Table 3 lists the details of the test population used in this study. For this work, we used cough sound data from a total of N1151 subjects (982 patients and 169 normal persons) to develop, validate and test our model. These patients were divided into two non-overlapping data sets: (1) training validation set (TrV) and (2) expected test set (PT). Patients are assigned to each set according to the order in which they visit the hospital.
Table 3: details of the total subject population used in the study
Training the validation dataset: for model training and validation, when we recruited 1011 subjects (852 patients and 159 normal) from two sites, we frozen the dataset; 600 patients and 134 normal persons.
Expected test data set: there were a total of 130 patients all from the JHC recording site, 10 normal subject records (9 from JHC, 1 from PMH).
Subjects in the data set were divided into a series of diagnostic subgroups according to clinical judgment and negotiation with a clinical team: normal group (Nr), primary upper respiratory tract infection group (U), laryngitis group (C), asthma/reactive airway disease group (a), clinical pneumonia group (P), bronchiolitis group (Bo), bronchitis group (Bc).
B. Training verifies the performance of the feature model on the data set.
Of the 1011 subjects, 725 subjects (602 patients, 123 normal subjects) were finally used for model training and validation.
TABLE 4 subgroups of diagnostic diseases
Table 4 shows the number of patients for each disease subgroup used to train the validation model.
LR models of normal versus disease: first, we explored the performance of the LR signature model in classifying cough in normal subjects and any disease subgroup subjects. Table 5 shows the leave-one-validation results of this exploration.
Table 5: leave-one-out validation results when classifying normal and disease coughs using the characteristic LR model.
As can be seen from table 5, all LR models were able to separate normal and disease coughs with very high accuracy after feature selection.
LR model among disease subgroups: our next goal was to explore the performance of the LR signature model in classifying coughs of two different disease groups. This exploration will help determine the extent to which the LR model captures the disease features. Table 6(A) shows the leave-one-out validation results of this search when all features were used for model training.
TABLE 6(A) leave-one-out validation Performance of the feature LR model when all features were used for model training
And Table 6(B) -spread-presenting results after feature selection
TABLE 6(B) leave-one-out validation Performance of the feature LR model when selected features are used for model training
According to table 6(B), most LR signature models achieved moderate to high accuracy (70-90%) in separating coughs from both categories after signature selection. The highest accuracy is achieved in identifying laryngitis or bronchiolitis cough from any other disease or group of diseases. The least accurate LR characteristic models were pneumonia and bronchitis and pURTI (accuracy-65%).
Performance of softmax model on training validation dataset
TABLE 7(A) leave-one-out validation Performance of the feature LR model when all features were used for model training
The results in section 3(B) indicate that the LR signature model is quite successful in capturing disease-specific signatures. Using these features in section 2(C), step 3, we trained the softmax neural network model to isolate the target disease cough from other diseases. Then using the cough index and applying the optimal threshold, we achieve the final goal of classifying the disease at the patient level.
Table 7(B) shows leave-one-out validation results of using the softmax-ANN model to isolate one disease from other diseases. According to the results, all models can predict target diseases except pneumonia, and have high sensitivity and specificity. The best validated results were obtained for the laryngitis model with a sensitivity of 100% and a specificity of 96% (using the cough plus simple clinical symptoms model), or with a sensitivity of 95% and a specificity of 92% (cough only model). The second best result is a model of bronchiolitis, followed by primary upper respiratory infection, asthma and a model of LRTD. All models trained using cough plus simple clinical features were significantly better than models trained using only cough features.
TABLE 7(B) leave-one-out validation results for classifying patients using the softmax ANN model
Leaving a cross-validation process will generate N softmax models. Where N is the number of patients in the TrV data set. From the N models, we select again an optimal model according to the k-means clustering algorithm. Table 8 shows the performance of the selected ANN models.
Table 8: performance of selected softmax ANN models on training-validation datasets
As can be seen from tables 7(B) and 8, the laryngitis, bronchiolitis, URTI and asthma models are very sensitive and specific. On the other hand, the pneumonia model has moderate specificity and high sensitivity. Therefore, we hypothesized that other disease models could be used as post-screeners in the pneumonia model to screen false positive cases and improve their specificity. To test this hypothesis, the inventors applied in sequence the asthma, primary upper respiratory infection, laryngitis and bronchiolitis models as post-screeners for the pneumonia model. To avoid truly positive cases of pneumonia from being screened out as other diseases, we used a threshold value for the cough index applied to the screening model. This threshold indicates how confidently the screening model indicates that the subject is not pneumonia. The screening threshold is optimized using the training validation dataset.
Table 9: results of the post-screener post-softmax pneumonia model were applied.
Table 9 (above) shows the results of the screener after sequential application of different diseases to the pneumonia model. It can be seen that the specificity of the ANN pneumonia model is increased with less loss of sensitivity. The sensitivity drop for the cough only model was-9%, and for the cough plus simple clinical characteristics model was-4%. The increase in specificity was 18% for the cough model alone and 16% for the cough plus simple clinical signature model. Although not critical, post-screeners and pre-screeners may be used in embodiments of the present invention.
Motivated by the positive impact of post-screeners on pneumonia models, we explored their application in other disease models. Our analysis shows that the specific performance of the bronchiolitis and LRTD models can be improved by 1-3% using the laryngitis screener. There was no improvement in the performance of the asthma and primary upper respiratory infection models. Since the performance of the laryngitis model is already very high, no screener was attempted. Table 10 shows the results of this study.
Table 10: results of applying post-screener to other disease models
Performance of softmax model on expected test data set
FIG. 3 shows a flow chart 300 of the diagnostic algorithm after all training is complete, all parameters are fixed, and the selected model is ready for prospective testing. These models were tested on an expected data set that was completely independent of the training validation data set. Except pneumonia, all models achieve high accuracy in predicting target diseases. In all models, laryngitis and bronchiolitis were best in all prospective study lists, with sensitivity and specificity in the range of 86% -100%, for cough only and cough plus simple clinical feature models. The asthma/RAD model only performed moderately on the cough signature, but with a simple clinical signature added to the cough signature, it was significantly improved (sensitivity-93%, specificity-90%). For detailed information on the pre-filter and post-filter, see appendix E.
1. Appendix A
1. Clinical pneumonia
Us clinical pneumonia case definition-data sets from JHC and PMH were labeled.
WHO Radioactive Primary Final pneumonia (PEP)
Us WHO radiolucent pneumonia case definition-used to label X-ray datasets from JHC and PMH.
3. Laryngitis (laryngitis)
Us clinical laryngitis case definition-used to label data sets from JHC and PMH.
4. Bronchiolitis
Us bronchiolitis case definition-used to label data sets from JHC and PMH.
5. Asthma (A)/Reactive Airway Disease (RAD) -A/RAD
Us a/RAD case definition-used to label data sets from JHC and PMH.
6. Bronchitis
Us bronchitis case definition-used to label data sets from JHC and PMH.
7. Upper respiratory tract infection
Us upper respiratory infection case definition-for labeling data sets from JHC and PMH.
8. Lower Respiratory Tract Diseases (LRTD)
US LRTD case definition-dataset for labeling JHC and PMH
2. Appendix B
Calculating features from coughs
Cough characteristics
Our method requires that a number of mathematical features be calculated from the cough sounds. This section describes that we derive from each sub-segment x of the recorded cough sound xiI is the calculated characteristic of 1, 2, 3.
i) Bispectral frequency cepstral coefficients (BFCC, 24 features in total; 8 per part of the cough segment) -the third-order spectrum of the signal is called bispectrum [3]. Unlike the power spectrum (second order spectrum based on autocorrelation), the bispectrum retains fourier phase information. From (6), segment x can be estimatediBispectrum B ofxi(ω1,ω2),
Wherein W (τ)1,τ2) Is a bispectrum window function, as used herein, the minimum bispectrum deviation supremum window, Cxi(τ1,τ2) Is x estimated by (2)iOf order 3,
In (7), Q is the length of the third order correlation lag under consideration, and xiIs a zero mean signal.
Bispectrum is a two-dimensional signal. However, it can be shown that for linear signals, except for slices parallel to the axis: any one-dimensional bispectrum slice other than ω 1-0, ω 2-0, and
Then apply the filter operator to the diagonal slice P (ω) and we use (8) to compute the bispectral frequency cepstral coefficients.
In (8), ζ represents a filter operator, where
table 11: the filter used to calculate the BFCC coefficients in equation (8) has a lower cutoff and a high cutoff.
ii) non-Gaussian score (NGS, 3 features total, 1 per part of cough segment) -NGS score is data xiA numerical measure of the non-gaussian nature of a given segment. A visual measure of the Gaussian of a set of data can be obtained using a normal probability map, with the NGS score being based on a regression scoreAnalytical non-gaussian quantification methods. We estimated the NGS score using (9), where p and q represent reference normal data and analytical data (x)i) Normal probability map of (2). The notation N is the number of data points used in the probability map.
iii) formant frequencies (total of 12 features, 4 per part of the cough segment) — in speech analysis, the Formant Frequencies (FF) are referred to as vocal tract resonances. In cough analysis, it is reasonably expected that the resonances of the entire airway that contribute to the generation of cough sounds will be manifested in the formant structure. A typical example is wheezing. The presence of mucus also changes the acoustic properties of the airway. We include the first four formants (F1, F2, F3, F4) in our candidate feature set. We passed the selection of cough fragment xiTo calculate F1-F4 from the peaks of the Linear Predictive Coding (LPC) spectrum. In this work, we used a 14 th order LPC model whose parameters were determined by the Levinson-Durbin (Levinson-Durbin) recursive process.
iv) log energy (LogE, 3 features, 1 per section of cough segment) -use (10) to calculate each sub-segment xiLogarithmic energy of (d):
in (4), it is an arbitrarily small normal number added to prevent the logarithm of 0 from being calculated unintentionally.
v) zero crossings (Zcr, 3 features, 1 per section of cough segment) — counting each sub-segment xiZero crossing number of (c).
vi) kurtosis (Kurt, 3 features total, 1 per part of the cough fragment) -kurtosis is xiA measure of how well the peaks of the probability density distribution are. It is xiCan be calculated by (11), where μ and σ denote x, respectivelyiAverage and standard deviation of.
vii) Mel frequency cepstral coefficients (MFCCs, 93 features total, 31 per part of the cough segment) -MFCCs have found widespread use in speech recognition systems. MFCC provides some resilience to non-speech variants in speech signals. They also provide orthogonal features to facilitate training of the classifier. The computation of the MFCC includes estimation of the short-term power spectrum, mapping of the Mel frequency score, and computation of the cepstral coefficients. In our work, we contained 31 MFCC coefficients in our feature set.
viii) shannon entropy (ShE, 3 features total, 1 per part of cough fraction): the cough sound is a complex signal representing the contribution of respiratory tract substructures. Some of these components have a pseudo-periodic structure, while others have randomness. In this work, we compute shannon entropy to capture these features. Each sub-segment x is calculated by (12)iShannon entropy (ShE).
ix) wavelet features (WvL, 13 features per cough): our previous studies have shown the usefulness of wavelet features of cough sounds in pneumonia diagnostics [ see ieee trans paper ]. For this work, we calculated 13 wavelet features from each cough segment. For details see [2 ].
3. Appendix C
1. Wheeze signal generator
Childhood wheezing is a common symptom of many respiratory diseases. Wheezing is defined as a high pitched whistle produced during breathing. Wheezing is often associated with asthma, but is also present in other respiratory diseases such as bronchiolitis, bronchitis, pneumonia, cystic fibrosis, and foreign body inhalation. It is often used for differential diagnosis and isolation of lower respiratory tract diseases and upper respiratory tract infections. See appendix a for more details: case definition.
The presence or absence of wheezing is a key decision node in the clinical decision tree of clinical community practice. However, clinical testing for this is not always a simple task. Wheezing is an unstable phenomenon and a secondary effect of the underlying changes in physiology/pathology. The ability of the physician to detect wheezing at a particular examination time depends on a number of factors, including the presence of wheezing at that time, and the clinician placing the stethoscope in the correct position over the lungs, the wheezing sound being generated with sufficient intensity to engender energy losses in the transmission from the lungs to the surface of the torso, and the clinician being able to perceive the sound and have the skill to detect it. The underlying physiological cause of interest is airway narrowing due to various causes, and wheezing is an alternative to this phenomenon. In some cases, wheezing may not occur, even if severe disease, because the severity of the disease causes airflow to be restricted (e.g., "rested chest" in severe asthma/RAD).
To understand the severity of wheezing, clinicians defined a number of different versions of Wheezing Severity Score (WSS). Our clinical partner version uses three different sub-scores to calculate WSS. These are: the presence of wheezing and its respiratory phase of occurrence, respiratory rate, use of auxiliary muscles. We have developed a feature to capture WSS using only coughing or increasing simple symptoms that parents can observe. Our WSS model was trained with continuous feature scores (LRM output) varying between 0-1, with the goal of separating high WSS (5-9) and low WSS scores (0, 1).
2. Lung function characteristic generator
Laboratory techniques of lung function, particularly spirometry, when available, can be used for the definitive diagnosis of certain respiratory diseases, such as asthma and Chronic Obstructive Pulmonary Disease (COPD). Spirometry provides numerical measurements, such as FEV1 and FVC. Using the coughs collected at spirometry, we established a model of characteristics based on coughs, such as: high FEV1 and low FEV1, high FEV1/FVC and
4. Appendix D
Clinical symptom-based features and diagnostic models
Clinicians are highly dependent on clinical symptoms (either they observe or parents report) when diagnosing certain respiratory diseases. Previously we have only studied and modeled on clinical symptoms to determine the best clinical profile for diagnosing pneumonia (indonesian study). We also investigated what happens when a small amount of cough is added to the clinical symptoms model.
We have established a clinical symptoms-only model (using parental reportable symptoms) to diagnose respiratory disease according to the structure shown in panel B. In one particular embodiment, symptoms such as fever, wheezing, runny nose, and age, gender are used to construct features based on the LRM model. The LRM model converts the classification symptoms into a continuous output at the feature generator level, which is then classified at the classifier block level. In other embodiments, other classifier schemes (e.g., ANN) may be used for the same purpose. We can also add cough to the clinical symptoms model to improve performance as we do in the clinical symptoms submitted by parents in the indonesia study. Furthermore, we can use clinical symptom features to improve classification using deep learning methods.
The process of designing a neural deep learning structure is divided into two phases, which are now discussed:
first stage-layer-by-layer training-in the first stage, we train three types of neural networks separately to accomplish a specific task.
First stage-neural
Stage 1-neural net type 2. Eigenneural network (SgNN): the output of the encoder is used to train a feed-forward neural network to generate the features described in the section "feature generation block". The feature ANN has no hidden layer and softmax neurons in the output layer. The advantage of using softmax neurons is that the probability function of the output, like the LR model, varies between 0 and 1, but now the mapping of input to output is non-linear. Fig. 6 shows an example of a characteristic neural net. The input to the network comes from the encoder. This neural network is trained to produce disease-specific features. In all 26 signature neural nets, all the signature models listed in table 1 were trained. Table 14 gives one-out validation results for classifying normal and disease coughs using the characteristic neural net model. Table 15 lists the leave-one-out validation performance of the feature neural net model.
Stage 1-neural net type 3. Classification neural network (CaNN): the output of the characteristic neural net is then used to train the final classification Softmax neural net. The neural network is similar to those described with reference to the
In a preferred embodiment, the inventors trained a single softmax-ANN model to identify all subgroups of target diseases: bronchiolitis, laryngitis, asthma/RAD, pneumonia, and lower respiratory tract disease. The characteristic neural network for training CaNN is
In the output layer, the k-dimension is set to 7 neurons, each neuron corresponding to a disease subgroup, bronchiolitis, asthma/RAD, pneumonia, bronchitis, pURTI, laryngitis and normal.Stage 2-the fine tuning stage in this stage, the separately trained neural networks from the first stage are connected together to create a stacked Deep Neural Network (DNN) as shown in fig. 7. Fig. 7 shows DNN, where the first layer represents a feature-encoding neural network, the second layer represents a feature neural network, and the third output layer represents a classification neural network.
The DNN is then retrained for a limited training period and the DNN network parameters are fine-tuned using the training validation dataset. The fine-tuning of the DNN is done according to leave-one-verify techniques.
Table 16 gives the leave-one-out validation results for classifying patients using the DNN model.
5. Appendix E
Pre-and post-screener
Another embodiment of the present invention will now be described with reference to the block diagram of the diagnostic model 300 shown on two pages in fig. 4 and 5. The diagnostic model 300 includes the disease feature generation block 114 and the
i) Prescreener block 111
The function of the pre-filter block 111 is to screen out subjects who are not intended to be analyzed by the main model.
Consider, for example, the case where: the task of the master model is defined as diagnosing a particular disease, such as bronchiolitis, from a mixture of other given diseases in a population of subjects visiting a medical institution. The result of the overall diagnostic algorithm is: "yes/no bronchiolitis? No or "is bronchiolitis yes/no? Is. In this case, the prescreener 111 may be designed to separate normal subjects from bronchiolitis cases and report the results as "yes/no for bronchiolitis? No case need be sent to the main model for further analysis.
Each of the screener models 4a, …, 4n in the screening box 111 has a decision threshold set high to ensure that people with actual disease targeted in the master model are not mistakenly screened out from further analysis by obtaining a no-disease label.
ii) a post-screener block 121
The function of the post-filter block 121 is to improve the diagnostic performance of the main model by determining the detection and correction of dominant false positives of the main model.
As an example, consider the case where the task of the master model is defined as diagnosing a particular disease, such as bronchiolitis, from a mixture of other given diseases in a population of subjects visiting a medical institution. The result of the overall diagnostic algorithm is: "yes/no bronchiolitis? No or "is bronchiolitis yes/no? Is. We hypothesized that in our primary model, laryngitis patients presented as a dominant false positive group. That is, in this group: "yes/no bronchiolitis? By, we found that a significant number of subjects were clinically diagnosed with laryngitis. In this scenario, our approach is to build a post-filter model: { laryngitis vs bronchiolitis }, which was trained to select laryngitis patients from a mix of laryngitis and bronchiolitis subjects. Then we will use our { laryngitis vs bronchiolitis } model to treat "bronchiolitis yes/no? Is "group, and moves detected laryngitis patients to" bronchiolitis yes/no? No side.
Multiple post-screener models 6a, …, 6m may be applied to a given master model, as needed and efficient. It may also be useful or necessary to have no post-screeners in a given master model. When needed, the screener model is used conservatively, with the decision threshold set high to ensure that people with the actual disease targeted in the main model do not erroneously shift to the other side of the diagnostic decision.
6. Appendix F
Feature reduction in model development
The diagnostic models discussed herein were developed using a cross-validation (CV) method on existing clinical data sets. The available data were subjected to K-fold cross-validation (K10) and leave-one-out-validation (LOOV) methods. Both methods have their own advantages and disadvantages. 10-fold cross validation has a tendency to result in lower variance and higher estimated bias for the model of the new dataset (generalization performance on previously unseen datasets); the LOOV has a tendency to result in higher variance and lower bias for the model of the new data set.
Because of the small size of the available data sets, it is generally preferred that the LOOV method be used more than 10-fold in developing diagnostic models. In order to compensate for the higher generalized variance in the model, a feature reduction method is proposed with the goal of making the built model as small as possible, thereby reducing the risk of model overfitting. The procedure is as follows.
Feature optimization/reduction is a technique to select a subset of relevant features to build a robust classifier. Optimal feature selection requires an exhaustive search of all possible feature subsets. However, this is impractical for a large number of features that we use as candidate features. Therefore, an alternative approach based on p-values is employed to determine salient features during the feature model design phase using Logistic Regression Models (LRMs). In LRM design, a p-value is calculated for each feature and indicates the importance of the feature to the model. The important features have a low p-value and this property of the LRM is used to determine an optimal combination of features that facilitates classification in the training phase model.
The method taken includes calculating an average P-value of all features across the entire data set and then selecting the average P-value to be less than a threshold PminThe characteristics of (1). The specific method is described below.
1. Let FN=[F1,F2,F3,…,FN]Representing an initial set of all N features.
2. Using FNAnd a Logistic Regression Model (LRM) is trained following the leave-one-out (or K-fold) method. The average p-value of all features is calculated.
PN=[Pf1,Pf2,Pf3,…,Pfn](1)
In (1), PNRepresenting a set of averages associated with the initial set of all N features.
3. By choosing an average P-value less than Pths=POIs characterized by FNCreating a new subset of features F'N。
4. According to the leave-one-to-verify (or K-fold) method, F 'is used'NThe LRM is trained on the feature set. Calculating F'NAverage p-value of the feature set.
5.By choosing an average P-value less than PthsIs characterized by being from F'NCreating a new feature subset F "N。
6. If F "NIs equal to F'NI.e. F'NIs less than PthsThen P is changed using the formula (2)ths
Pths=Pths-c(2)
Wherein c < Pths
7. Repeat step 6 until F "NAggregate size less than F'NAnd (4) collecting.
8. Will PthsIs reset to the initial value and steps 4-7 are repeated until F "NFeature set size less than Fmin。
9. For each subset of selected features, model performance (sensitivity, specificity and kappa number) is calculated. To select the best subset of features, we follow the following scheme.
a. The subset of features with the largest kappa number in the model performance is selected. Let the subset be Fa。
b. Now using FaThe feature subset identifies all feature subsets whose performance (in terms of sensitivity/specificity) is within q% of the model performance.
c. From the identified pool of feature subsets, a subset is selected that has a minimum size, but is larger than Z and smaller than FaThe size of (2). If this condition is not met, then F is selectedaAs an optimal feature set.
In one particular embodiment, the following parameter values are used with the algorithm described above. Cough model only FNThe size is 157 (i.e., 157 input features that begin using the LRM). Cough plus clinical symptom model FNThe size is 157+ depending on the number of clinical features used.
Pths(in step 3) ═ PO=0.20
c (in step 6) is 0.001
Fmin (at step 8) ═ 10
Z (in step 9) ═ 10
7. In a particular embodiment, the following parameter values and the above algorithm are used. The results obtained in this way are shown in tables 12 and 13 below.
TABLE 12 leave-one-verified performance of the feature LR model after feature optimization. Only the cough signature.
TABLE 13 leave-one-verified performance of the feature LR model after feature optimization. Cough plus simple clinical features
TABLE 14 leave-on validation results for classifying normal and disease coughs using the characteristic neural net model
TABLE 15 leave-one-verification Performance of the characteristic neural network model
TABLE 16 leave-one-out validation results when classifying patients using DNN model
Reference to the literature
The following documents are incorporated herein by reference in their entirety.
1.Abeyratne,U.R.,et al.,Cough sound analysis can rapidly diagnosechildhood pneumonia.Annals of biomedical engineering,2013.41(11):p.2448-2462.
2.Kosasih,K.,et al.,Wavelet augmented cough analysis for rapidchildhood pneumonia diagnosis.IEEE Transactions on Biomedical Engineering,2015.62(4):p.1185-1194.
3.Abeyratne,U.Blind reconstruction of non-minimum-phase systems from1-D oblique slices of bispectrum.1999.IET.
4.Hinton GE,Salakhutdinov RR.Reducing the dimensionality of data withneural networks.science.2006Jul 28;313(5786):504-7.
In compliance with the statute, the invention has been described in language more or less specific as to structural or methodical features. The word "comprising" and variations thereof, such as "comprises" and "comprising", are used throughout in an inclusive sense without precluding any additional features. It is to be understood that the invention is not limited to the specific features shown or described, since the means herein described comprise preferred forms of putting the invention into effect. The invention is, therefore, claimed in any of its forms or modifications within the proper scope of the appended claims appropriately interpreted by those skilled in the art.
In the specification and claims (if any), the terms "substantially" or "about" will be understood to mean a value not limited to the range specified by the term, unless the context requires otherwise.
Features, integers, characteristics, compounds, chemical moieties or groups described in conjunction with a particular aspect, embodiment or example of the invention are to be understood to be applicable to any other aspect, embodiment or example described herein unless incompatible therewith.
Any embodiments of the present invention are intended to be illustrative, and not restrictive. It is therefore to be understood that various other adaptations and modifications may be made to any of the described embodiments without departing from the spirit and scope of the invention.
- 上一篇:一种医用注射器针头装配设备
- 下一篇:用于检测二氧化碳的装置