阅读说明：本技术 基于概率神经网络模型的心音分析系统 (Heart sound analysis system based on probabilistic neural network model ) 是由 史沛然 李秋霖 完晓妍 于 2018-08-13 设计创作，主要内容包括：一种基于概率神经网络模型的心音分析系统。这个系统由心音采集处理模块和心音智能分析模块构成,硬件部分为心音采集电路单元,软件部分为心音处理单元、心音自动诊断单元以及数据显示单元。本系统通过对心音信号的采集、处理和智能分析可以实现常见心脏疾病的自动诊断,给人们提供健康信息和健康建议,也可作为医院辅助诊断设备。(A heart sound analysis system based on a probabilistic neural network model. The system consists of a heart sound acquisition and processing module and a heart sound intelligent analysis module, wherein the hardware part is a heart sound acquisition circuit unit, and the software part is a heart sound processing unit, a heart sound automatic diagnosis unit and a data display unit. The system can realize automatic diagnosis of common heart diseases through acquisition, processing and intelligent analysis of heart sound signals, provides health information and health advice for people, and can also be used as auxiliary diagnosis equipment of hospitals.)

1. A heart sound analysis system based on a probabilistic neural network model is characterized by comprising an original heart sound acquisition and processing module and a heart sound intelligent analysis module; wherein the content of the first and second substances,

the original heart sound acquisition and processing module is used for acquiring and preprocessing heart sound signals;

the heart sound intelligent analysis module is used for inputting the preprocessed heart sound into the probabilistic neural network model and realizing automatic diagnosis of certain common heart diseases by identifying the waveform of the heart sound.

2. The probabilistic neural network model-based heart sound analysis system of claim 1, wherein the raw heart sound collection, processing and analysis module specifically comprises:

the heart sound acquisition unit is used for acquiring heart sounds of the heart to be detected;

the heart sound normalization processing unit is used for carrying out normalization processing of denoising, truncation and envelope extraction on the collected heart sound signals;

and the heart sound characteristic extraction unit is used for extracting heart sound characteristic parameters for intelligent analysis of the heart sound.

3. The system of claim 1, wherein the intelligent analysis module comprises:

the automatic diagnosis unit based on the probabilistic neural network model brings the heart sound characteristic parameters into the probabilistic neural network model for realizing the diagnosis of the heart health condition and several common heart diseases;

the diagnosis result display unit is used for displaying the result of the automatic diagnosis to the user and giving a simple health suggestion;

and the health data storage unit is used for storing the measured heart sound information so as to be used when professional medical advice is sought.

4. A method of operating a probabilistic neural network model based heart sound analysis system as claimed in claim 1, comprising the steps of;

step 1: collecting heart sounds of a user;

step 2: inputting the collected heart sounds into a computer, carrying out preprocessing of noise elimination, truncation and envelope extraction, and extracting heart sound characteristic parameters;

and step 3: the heart sound characteristic parameters are brought into the probabilistic neural network model to obtain the health condition information of the heart, and several common heart diseases are automatically diagnosed;

and 4, step 4: the heart sound intelligent analysis result is displayed to the user, and a simple health suggestion is provided;

and 5: the analysis result is stored, so that the query and the use are convenient in the future.

Technical Field

The invention relates to the field of medical analysis, in particular to a system for processing and analyzing heart sounds.

Background

According to the statistics of the world heart alliance on death people worldwide, one third of the deaths are cardiovascular diseases, and 80% of the deaths from cardiovascular diseases are from middle-low income countries or regions. Heart disease has become a non-negligible "health killer". Therefore, effective measures for preventing and treating cardiovascular diseases need to be provided urgently.

Clinically, the diagnosis of heart and cardiovascular diseases is divided into two types, namely non-invasive and invasive, and the non-invasive diagnosis (including an impedance cardiogram, a cardiac apical pulsation cardiogram, an ultrasonic cardiogram and the like) has problems in the aspects of measurement accuracy, simplicity and stability, so that the wide application of the diagnosis is limited; invasive diagnosis (including cardiovascular radiography, floating catheter method and the like) has good effect, but belongs to invasive detection, has certain risk and high cost, and is not generally accepted yet. The electrocardiogram diagnosis still has the defect of diagnosis lag, for example, the electrocardiogram can be changed when the blockage reaches more than 70-75% in the diagnosis of coronary artery blockage diseases, and the heart sound signal can be changed when the blockage rate is 25%, which indicates that the heart sound contains abundant cardiovascular information, and the analysis of the heart sound signal has extremely important practical significance in the diagnosis and prevention of cardiovascular system diseases.

Today, the heart auscultation is still an important auxiliary diagnostic means in the explosive development of modern medical science and technology, especially heart diagnosis technology, and heart sound signals obtained by the heart auscultation contain a large amount of pathological information. The traditional auscultation technology excessively depends on the experience judgment of doctors, and the problems of judgment errors or low diagnosis efficiency and the like easily occur.

Disclosure of Invention

The invention aims to realize heart sound-based heart auxiliary diagnosis by utilizing a neural network model in machine learning so as to overcome the problems.

The purpose of the invention can be realized by the following technical scheme:

(1) and reading the heart sound recording by using an audio reading function in the audio software working platform.

(2) And (4) denoising the heart sound signal by using a threshold.

(3) And carrying out stage analysis on the heart sound signals to obtain truncated heart sound waveforms.

(4) And performing multi-scale wavelet decomposition on the preprocessed heart sound signals, and performing five-layer wavelet decomposition on the heart sound signals. The low-frequency coefficient obtained by wavelet decomposition and the high-frequency coefficients of the fourth layer and the fifth layer are omitted. The first layer of high-frequency coefficients represents frequency band information of normal heart sounds, and the second layer of high-frequency coefficients and the third layer of high-frequency coefficients represent frequency band information of possible heart murmurs.

(5) Normalizing the sampling data according to the following formula:

whereinIn order to be the original signal, the signal is transmitted,

is the maximum value of the amplitude in the time domain. The sound splitting condition can be observed in the preprocessed time domain waveform diagram of part of sound data.

(6) Segmenting data, wherein each 15ms is a segment, the interval is 5ms, according to the definition of average Shannon energy, normalized Shannon energy of the whole segment of heart sound signal is calculated by the segmented data, a first heart sound and a second heart sound are directly identified on a time domain graph of normalized Shannon energy distribution, further a cardiac cycle (hrt), a time delay (s1t, s2t) and an amplitude (I1, I2) of the first and second heart sounds are obtained, and finally, calculating a characteristic value of the heart sounds in the time domain comprises the following steps: heart rate per minute (hr), S1 time limit to heart rate ratio (S1 t/hr), S2 time limit to heart rate ratio (S2 t/hr), S1 amplitude to S2 amplitude ratio (I1/I2).

(7) The heart sound diagnosis model takes the selected characteristic value as the input of the model and carries out diagnosis by using a Probabilistic Neural Network (PNN).

The diagnosis process is divided into a training process and an actual application process. In the training process:

1 setting PNN network parameter initial value

And 2, inputting the training set into the PNN network for training, and adjusting parameters to obtain a diagnosis model.

And 3, inputting the test set into the network for testing.

And inputting a sample to be diagnosed in the diagnosis process, and finally obtaining an output diagnosis result.

Compared with the prior art, the invention has the following advantages:

1. wavelet transformation and normalization Shannon processing can comprehensively and accurately represent time domain characteristics of heart sounds.

2. The PNN model is high in building speed and simple in learning process, the number of hidden layer neurons can be determined according to samples, complex repeated training is not needed, and the model construction efficiency is high.

The diagnosis of PNN is based on minimum risk Bayes decision making, and can obtain more accurate result than the prior BP algorithm.

Drawings

FIG. 1 is a block diagram of an acquisition circuit system;

FIG. 2 is a flow chart of the algorithm of the present invention;

fig. 3 is a flow chart of probabilistic neural network analysis of heart sounds.

Detailed Description

The invention will be further described with reference to the accompanying drawings in which:

as shown in fig. 1, the collecting circuit of the present invention is an available collecting circuit, which uses a mechanical stethoscope head and an electret microphone as a signal collector and a sensor, and an analog signal processing part for amplification and filtering, and plays, stores and further processes the signals in an earphone, a single chip microcomputer and an upper computer.

Fig. 2 shows a specific algorithm flow of the analysis processing after the signal acquisition of the present invention.

In the heart sound signal preprocessing process, threshold denoising is firstly carried out on data, then five-layer wavelet decomposition is carried out on the heart sound signals, and because different wavelet decomposition coefficients represent signal information of different frequency bands, low-frequency coefficients, fourth-layer high-frequency coefficients and fifth-layer high-frequency coefficients are omitted. And preserving the high-frequency coefficients of the first layer, the second layer and the third layer in the decomposition result. The first layer of high-frequency coefficients are frequency band information of normal heart sounds, and the second layer of high-frequency coefficients and the third layer of high-frequency coefficients are frequency band information of possible heart murmurs.

The process of selecting the diagnostic characteristic value firstly carries out segmentation processing on the data in a time domain, and then carries out normalization processing on the sampled data to obtain the heart sound diagnostic characteristic values such as a cardiac cycle (hrt), a heart rate per minute (hr), an S1 time limit and heart rate ratio (S1 t/hr), an S2 time limit and heart rate ratio (S2 t/hr), an S1 amplitude value and an S2 amplitude value (I1/I2) and the like.

Fig. 3 shows a flow of analyzing heart sounds using a probabilistic neural network according to the present invention.

The first layer is an input layer which transmits the input heart sound characteristics to all the mode units, and the number of the neurons is equal to the length of the heart sound characteristic vector. The input heart sound characteristic vector is first with the coefficient omega_iAnd inputting the weighted result into a mode layer for calculation.

The second layer is a mode layer, and the number of the neurons of the mode layer is equal to the total number of the training samples. The probability of each neuron output in each type of heart sound is calculated in the mode layer, and the calculated probability is input into the summation layer for further analysis.

The third layer is a summation layer, the number of neurons in the summation layer is the same as that of the heart sound types, the weighted average is carried out on the output of the hidden neurons belonging to the same heart sound type in the mode layer, and the probability density function of each heart sound type is obtained.

And the fourth layer is an output layer, and the output layer selects the heart sound type with the maximum posterior probability as a diagnosis result by adopting Bayes classification rules according to the estimation of the probability of each type of input vectors.