Heart conduction health degree assessment method, medium and system

Abstract

The invention provides a heart conduction health degree assessment method, medium and system, and belongs to the technical field of electrical digital data processing.The heart conduction health degree assessment method comprises the steps that firstly, heart conduction signals of multiple testers and expert assessment results are collected; splitting the signals by using an empirical mode decomposition method to obtain a standard signal matrix and a non-standard signal matrix; then calculating the wavelet feature of the standard signal and the Fourier feature of the non-standard signal; based on the features, an improved convolutional neural network model is constructed, including a spatial pyramid pooling layer, an attention mechanism, residual connection and multi-scale feature fusion. After being trained, the model can be used for evaluating the health degree of a to-be-evaluated heart conduction signal. According to the method, the time-frequency characteristic of the signal is fully utilized, more accurate health condition assessment is realized through improvement of the deep learning model, and the problem that the overall health condition is difficult to comprehensively and objectively assess since most of the prior art only focuses on the local characteristic of the heart conduction system is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of electrical digital data processing, and in particular, relates to a method, medium and system for evaluating the health of cardiac conduction. Background Art

[0002] Cardiac conduction function is a key process to maintain normal heartbeat, and its abnormality can lead to various heart diseases. Existing cardiac health assessment mainly relies on the measurement and analysis of various physiological signals such as electrocardiogram (ECG), phonocardiogram (PCG), cardiac ultrasound (ECHO) and cardiac magnetic resonance (CMR). These signals can reflect the structure, function, physiology and pathological characteristics of the cardiac conduction system. The cardiac health assessment methods widely used in clinical practice mainly include the following:

[0003] 1. Diagnostic methods based on electrocardiograms. Electrocardiograms can record the changes in cardiac electrical activity, thereby reflecting the state of myocardial excitation and propagation. By analyzing the characteristics of electrocardiogram waveforms, such as P waves, QRS complexes, and T waves, common heart diseases such as arrhythmias, myocardial infarction, and myocardial hypertrophy can be diagnosed. However, electrocardiograms can only reflect the characteristics of myocardial electrical activity and cannot fully understand the overall health of the cardiac conduction system.

[0004] 2. Diagnostic methods based on phonocardiogram. The phonocardiogram records the acoustic signals generated by the opening and closing of each heart valve and hemodynamic changes. By analyzing the characteristics of the phonocardiogram waveform, such as the first heart sound S1, the second heart sound S2, and additional heart sounds such as murmurs, various heart valve diseases and hemodynamic abnormalities can be diagnosed. However, the phonocardiogram only reflects the local characteristics of the heart's mechanical activity, and it is difficult to fully evaluate the overall health level of the heart's conduction function.

[0005] 3. Diagnostic methods based on cardiac ultrasound. Cardiac ultrasound can dynamically observe the morphology and function of each heart chamber, such as ejection fraction, valve activity, etc. By analyzing the quantitative parameters of ultrasound images, diseases such as myocardial hypertrophy, heart valve disease, and myocardial infarction can be diagnosed. However, ultrasound imaging is greatly affected by factors such as patient size and operator experience, and it is difficult to fully evaluate the overall function of the cardiac conduction system.

[0006] 4. Diagnostic methods based on cardiac magnetic resonance. Cardiac magnetic resonance imaging can clearly display the morphological structure, perfusion status and tissue properties of the myocardium. By analyzing the quantitative parameters of magnetic resonance images, such as ventricular volume and myocardial fiber orientation, diseases such as cardiomyopathy, myocardial infarction and myocarditis can be diagnosed. However, cardiac magnetic resonance examinations are expensive and require professional operators, making them unsuitable for large-scale daily heart health monitoring.

[0007] In summary, although the existing cardiac health assessment technologies have played an important role in their respective application scenarios, most of them can only focus on the local characteristics of the cardiac conduction system, and it is difficult to comprehensively and objectively assess the overall health of the cardiac conduction system. At the same time, these methods have certain limitations in terms of operational difficulty, inspection cost, and frequency of use, and cannot meet the needs of daily cardiac health monitoring. Therefore, there is an urgent need to develop a new cardiac conduction health assessment method that can comprehensively utilize multiple physiological signal characteristics to comprehensively and accurately assess the overall conduction function of the heart. Summary of the invention

[0008] In view of this, the present invention provides a cardiac conduction health assessment method, medium and system, which can solve the technical problem that most existing technologies can only focus on local characteristics of the cardiac conduction system and are difficult to comprehensively and objectively assess the overall health status of the cardiac conduction system.

[0009] The present invention is achieved in that:

[0010] A first aspect of the present invention provides a method for evaluating cardiac conduction health, comprising the following steps:

[0011] S10, obtaining a plurality of cardiac conduction signals of a plurality of test persons and obtaining expert evaluation data of the cardiac conduction health of the test persons;

[0012] S20, performing decomposition calculation on each signal respectively to obtain a standard signal matrix and a non-standard signal matrix;

[0013] S30, according to a preset splitting threshold, determining a signal in the standard signal matrix whose similarity to the corresponding standard signal is higher than the splitting threshold;

[0014] S40, using a wavelet transform method to calculate the characteristics of the signals in the standard signal matrix to obtain standard characteristics;

[0015] S50, using a Fourier transform method to calculate the characteristics of the signal in the non-standard signal matrix to obtain non-standard characteristics;

[0016] S60, constructing a cardiac conduction health assessment model based on the standard features and the non-standard features;

[0017] S70, training the heart conduction health assessment model to obtain a trained heart conduction health assessment model;

[0018] S80, obtaining a cardiac conduction signal to be evaluated, inputting the cardiac conduction signal to be evaluated into the trained cardiac conduction health assessment model, obtaining a cardiac conduction health assessment result and outputting it.

[0019] On the basis of the above technical solution, a cardiac conduction health assessment method of the present invention can also be improved as follows:

[0020] The multiple signals conducted by the heart include electrocardiogram signals, heart sound signals, cardiac ultrasound signals and cardiac magnetic resonance signals.

[0021] Furthermore, the test persons include persons with different health conditions and patients with heart diseases.

[0022] Furthermore, the cardiac conduction health assessment model specifically adopts a convolutional neural network structure, and replaces the pooling layer with a spatial pyramid pooling layer, and adds an attention mechanism module, a residual connection module, and a multi-scale feature fusion module to obtain the cardiac conduction health assessment model. This improved neural network structure effectively improves the model's ability to process multimodal cardiac signals by introducing modules such as a spatial pyramid pooling layer, an attention mechanism, a residual connection, and a multi-scale feature fusion. The spatial pyramid pooling layer solves the limitation of the fixed length of the input signal and retains more spatial information; the attention mechanism can automatically identify important features and improve interpretability; the residual connection alleviates the gradient vanishing problem of the deep network and accelerates convergence; and the multi-scale feature fusion enhances the model's ability to express complex cardiac conduction patterns. These improvements work together to enable the model to better adapt to the temporal and multi-scale characteristics of cardiac signals, improve the detection accuracy of different types of cardiac conduction abnormalities, and enhance the generalization ability of the model and the reliability of the evaluation results.

[0023] Furthermore, this improved neural network structure processes standard features and non-standard features separately through a unique dual-channel architecture. The standard features utilize the good time-frequency localization characteristics obtained by wavelet transform, retain multi-scale information through the spatial pyramid pooling layer, and the attention mechanism highlights the key waveform and periodic features. The non-standard features utilize the complete frequency domain characteristics obtained by Fourier transform, detect abnormal frequency components through multi-scale feature fusion, and the attention mechanism locates abnormal waveforms. The features of the two channels remain independent in the deep feature extraction process, and finally dynamic weight allocation and effective integration are achieved through the multi-scale feature fusion module and attention mechanism. This processing method not only maintains the unique properties of the two types of features, but also realizes the complementary fusion of features, improves the sensitivity of abnormality detection, and provides a more comprehensive and reliable feature representation for cardiac conduction health assessment.

[0024] Furthermore, the decomposition calculation specifically uses an empirical mode decomposition method to decompose the signal.

[0025] Furthermore, the similarity adopts cosine similarity.

[0026] Furthermore, the step of obtaining the splitting threshold includes:

[0027] Step 1: Collect a large number of cardiac conduction signal samples from healthy people and patients with known heart diseases;

[0028] Step 2: pre-process the sample;

[0029] Step 3, calculating the similarity between the sample signal and the standard signal;

[0030] Step 4: Use the support vector machine algorithm to classify the similarities;

[0031] Step 5: Determine the optimal split threshold through cross-validation and ROC curve analysis.

[0032] Furthermore, the formula for the multivariate mixed decomposition calculation is expressed as follows:

[0033]

[0034] Where x(t) is the original signal; c i (t) is the i-th intrinsic mode function (IMF); r n (t) is the residual signal; n is the number of IMFs.

[0035] For each IMF, the calculation process is as follows:

[0036] h1(t)=x(t)-m1(t);

[0037] h k (t) = h k-1 (t)-m k (t);

[0038] In the formula, m k (t) is h k-1 (t) is the average value of the upper and lower envelopes; k is the number of iterations.

[0039] The formula of the wavelet transform method is as follows:

[0040]

[0041] Where W f (a, b) are wavelet transform coefficients; f(t) is the signal; ψ(t) is the wavelet function; a is the scale parameter; b is the translation parameter; * represents complex conjugate.

[0042] The formula of the Fourier transform method is as follows:

[0043]

[0044] Where F(ω) is the frequency spectrum of the signal f(t); ω is the angular frequency; j is the imaginary unit.

[0045] The convolutional layer calculation formula of the cardiac conduction health assessment model is as follows:

[0046]

[0047] In the formula, is the k-th feature map of the l-th layer; σ is the activation function; is the convolution kernel; For input; is the bias term; * indicates the convolution operation.

[0048] The calculation formula of the spatial pyramid pooling layer is as follows:

[0049] z l =concat[pool1(x l ),pool2(x l ),...,pool L (x l )];

[0050] In the formula, z l is the feature after pooling; pool i represents the i-th level pooling operation; L is the number of pyramid levels; concat() represents feature concatenation.

[0051] The calculation formula of the attention mechanism module is as follows:

[0052] α=softmax(W2tanh(W1x+b1)+b2);

[0053] y=α⊙x;

[0054] Where α is the attention weight; W1, W2 are weight matrices; b1, b2 are bias terms; x is the input feature; y is the weighted feature; ⊙ represents element-by-element multiplication.

[0055] The calculation formula of the residual connection module is as follows:

[0056] y=F(x,{W i})+x;

[0057] Where F represents the residual function; x is the input; {W i} is a weight set.

[0058] The calculation formula of the multi-scale feature fusion module is as follows:

[0059]

[0060] In the formula, y is the fused feature; w i is the weight coefficient; xi are features of different scales; upsample represents the upsampling operation; N is the number of feature scales.

[0061] Furthermore, the step of obtaining the splitting threshold specifically includes:

[0062] 1. Collect a large number of cardiac conduction signal samples from healthy people and patients with known heart diseases;

[0063] 2. Preprocess and standardize the samples. The mathematical expression of preprocessing is as follows:

[0064]

[0065] In the formula, x norm is the standardized signal; x is the original signal; μ is the mean of the signal; σ is the standard deviation of the signal.

[0066] 3. Calculate the similarity between the sample signal and the standard signal using cosine similarity, which is mathematically expressed as follows:

[0067]

[0068] Where A and B are sample signal and standard signal respectively; n is the dimension of the signal; A i and B i are the i-th components of A and B respectively.

[0069] 4. Use support vector machine (SVM) to classify similarities, and its decision function is as follows:

[0070]

[0071] In the formula, x is the input feature; x i is the support vector; y i is the category label; α i is the Lagrange multiplier; K(x i ,x) is the kernel function; b is the bias term; sign is the sign function.

[0072] 5. Determine the optimal split threshold through cross-validation and ROC curve analysis. The mathematical expression of the ROC curve is as follows:

[0073]

[0074]

[0075] Where TPR is the true positive rate; FPR is the false positive rate; TP is the number of true positives; FP is the number of false positives; TN is the number of true negatives; and FN is the number of false negatives.

[0076] The selection of the optimal splitting threshold is based on Youden's Index:

[0077] J=max(TPR+TNR-1)=max(TPR-FPR);

[0078] Where J is the Youden index; TNR is the true negative rate, which is equal to 1-FPR.

[0079] By maximizing the Youden index, the optimal splitting threshold T can be obtained opt :

[0080] T opt = argmax T (TPR(T)-FPR(T));

[0081] Where T is the candidate threshold; TPR(T) and FPR(T) are the true positive rate and false positive rate at the threshold T, respectively.

[0082] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the above-mentioned method for assessing cardiac conduction health.

[0083] A third aspect of the present invention provides a cardiac conduction health assessment system, which includes the above-mentioned computer-readable storage medium.

[0084] Compared with the prior art, the beneficial effects of the cardiac conduction health assessment method, medium and system provided by the present invention are:

[0085] 1. Comprehensive signal acquisition. A variety of physiological signals are collected to fully reflect the health of the cardiac conduction system.

[0086] 2. In-depth feature extraction. With the help of EMD decomposition and deep learning technology, rich time-frequency features can be mined.

[0087] 3. High evaluation accuracy. Deep learning training based on a large number of samples can accurately evaluate the health of cardiac conduction.

[0088] In summary, the cardiac conduction health assessment method proposed in the present invention breaks through the limitations of the existing technology and solves the technical problem that most of the existing technologies can only focus on the local characteristics of the cardiac conduction system and are difficult to comprehensively and objectively assess the overall health status of the cardiac conduction system. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 A flow chart of the method provided by the present invention;

[0090] Figure 2This is the age distribution diagram of subjects in different health status;

[0091] Figure 3 This is the EMD decomposition result diagram of the electrocardiogram signal;

[0092] Figure 4 This is a comparison diagram of wavelet time-frequency diagram;

[0093] Figure 5 Model performance graph for cardiac conduction fitness assessment. DETAILED DESCRIPTION

[0094] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0095] like Figure 1 FIG. 1 is a flow chart of a method for evaluating cardiac conduction health provided by the first aspect of the present invention. The method comprises the following steps:

[0096] S10, obtaining multiple cardiac conduction signals of multiple test persons and obtaining expert evaluation data of the cardiac conduction health of the test persons;

[0097] S20, performing decomposition calculation on each signal respectively to obtain a standard signal matrix and a non-standard signal matrix;

[0098] S30, according to a preset splitting threshold, determining a signal in the standard signal matrix whose similarity to the corresponding standard signal is higher than the splitting threshold;

[0099] S40, using a wavelet transform method to calculate the characteristics of the signal in the standard signal matrix to obtain standard characteristics;

[0100] S50, using a Fourier transform method to calculate the characteristics of the signal in the non-standard signal matrix to obtain non-standard characteristics;

[0101] S60. Construct a cardiac conduction health assessment model based on standard features and non-standard features;

[0102] S70, training the heart conduction health assessment model to obtain a trained heart conduction health assessment model;

[0103] S80, obtaining a cardiac conduction signal to be evaluated, inputting the cardiac conduction signal to be evaluated into a trained cardiac conduction health assessment model, obtaining a cardiac conduction health assessment result and outputting it.

[0104] The specific implementation methods of the above steps are described in detail below:

[0105] Step S10: Acquire multiple cardiac conduction signals of multiple test persons:

[0106] The purpose of this step is to collect a variety of cardiac conduction-related physiological signals including electrocardiogram, heart sound, cardiac ultrasound and cardiac magnetic resonance, etc. The specific implementation is as follows:

[0107] 1. Select subjects with different health conditions, including normal people and patients with known heart diseases. This can cover both healthy and unhealthy heart conduction characteristics.

[0108] 2. Collect various physiological signals from each subject, including electrocardiogram, heart sound, cardiac ultrasound and cardiac magnetic resonance, etc. These signals reflect the structure, function, physiology and pathology of the cardiac conduction system.

[0109] 3. At the same time, the expert evaluation data of each subject's cardiac conduction health was recorded. The experts gave an evaluation of the cardiac conduction health status based on the clinical diagnosis and examination results.

[0110] 4. This step can obtain a rich sample library of cardiac conduction signals, laying the foundation for subsequent model training and performance evaluation.

[0111] Step S20, perform multivariate mixed decomposition calculation on each signal:

[0112] The purpose of this step is to decompose and analyze the obtained cardiac conduction signal and extract the standard signal and non-standard signal components. This is achieved using the empirical mode decomposition (EMD) algorithm. The specific steps are as follows:

[0113] 1. Apply the EMD algorithm to decompose each original cardiac conduction signal x(t). The mathematical expression of EMD is:

[0114]

[0115] In the formula, c i (t) represents the i-th intrinsic mode function (IMF), r n (t) is the residual signal, and n is the number of IMFs.

[0116] 2.IMFc i The calculation formula of (t) is as follows:

[0117] h1(t)=X(t)-m1(t);

[0118] h k (t) = h k-1 (t)-m k (t);

[0119] Among them, m k (t) is h k-1The average value of the upper and lower envelopes of (t), k is the number of iterations.

[0120] 3. Through the above EMD decomposition process, the original cardiac conduction signal x(t) can be decomposed into a series of IMFc i (t) and the residual term r n (t). Where IMFc i (t) represents the standard signal component, r n (t) represents the non-standard signal component.

[0121] Step S30: Determine the standard signal matrix:

[0122] The purpose of this step is to filter out the part that is most similar to the standard signal from the signal matrix obtained by EMD decomposition. The specific steps are as follows:

[0123] 1. Set a predetermined similarity threshold T split The threshold can be determined by adaptive learning from a large number of samples through a machine learning algorithm (such as SVM).

[0124] 2. For each IMFc i (t) is calculated with the similarity of the standard signal. Here, cosine similarity is used for measurement, and its formula is:

[0125]

[0126] In the formula, A and B are sample signal and standard signal respectively, and n is the dimension of the signal.

[0127] 3. Compare each IMFc i The similarity between (t) and the preset threshold T split When the similarity is higher than the threshold, the IMF is classified as a standard signal matrix; otherwise, it is classified as a non-standard signal matrix.

[0128] Through the processing of this step, the standard signal component most similar to the normal healthy state can be separated from the original complex cardiac conduction signal.

[0129] Step S40, calculating the characteristics of the standard signal:

[0130] The purpose of this step is to extract the time-frequency characteristics of the signal components selected from the standard signal matrix.

[0131] Wavelet transform method is used for feature extraction. The specific steps are as follows:

[0132] 1. Apply wavelet transform to each signal f(t) in the standard signal matrix and calculate the wavelet transform coefficient W f (a,b):

[0133]

[0134] Where ψ(t) is the wavelet function, a is the scale parameter, and b is the translation parameter.

[0135] 2. The time-frequency characteristics of the standard signal can be obtained through wavelet transform. These characteristics can reflect the regularity and stability of the heart conduction signal under the standard healthy state.

[0136] Step S50: Calculate the characteristics of the non-standard signal:

[0137] The purpose of this step is to extract the spectral features of the non-standard signal components filtered out from the original signal matrix. The Fourier transform method is used for feature extraction. The specific steps are as follows:

[0138] 1. Apply Fourier transform to each signal f(t) in the non-standard signal matrix and calculate the spectrum F(ω):

[0139]

[0140] Where ω is the angular frequency and j is the imaginary unit.

[0141] 2. The spectrum characteristics of non-standard signals can be obtained through Fourier transform. These characteristics can reflect the spectrum characteristics of signals when abnormal cardiac conduction exists.

[0142] Step S60: construct a cardiac conduction health assessment model:

[0143] The purpose of this step is to build a deep learning model that can evaluate the health status of cardiac conduction based on the standard features and non-standard features extracted above. The specific implementation is as follows:

[0144] 1. Use convolutional neural network (CNN) as the basic model structure. CNN can effectively extract the spatiotemporal features of signals.

[0145] 2. Replace the pooling layer of CNN with the spatial pyramid pooling layer (SPP). SPP can adaptively extract multi-scale features and enhance the model's ability to learn features of different scales. Its calculation formula is:

[0146] z l =concat[pool1(x l ),pool2(x l ),...,pool L (x l )];

[0147] In the formula, z l is the feature after pooling, pool i represents the i-th level pooling operation, and L is the number of pyramid levels.

[0148] 3. Add an attention mechanism module to the CNN main network. The attention mechanism can adaptively assign weights to different features and highlight key features. Its calculation formula is:

[0149] α=softmax(W2tanh(W1x+b1)+b2);

[0150] y=α⊙x;

[0151] Where α is the attention weight, W1, W2 are weight matrices, and b1, b2 are bias terms.

[0152] 4. Apply the residual connection module to enhance the gradient flow and feature learning ability of the model. The calculation formula is:

[0153] y=F(x,{W i})+x;

[0154] Where F represents the residual function, x is the input, {W i} is a weight set.

[0155] 5. Use the multi-scale feature fusion module to integrate features at different levels. The calculation formula is:

[0156]

[0157] In the formula, y is the fused feature, w i is the weight coefficient, x i are features of different scales, and upsample() represents an upsampling operation.

[0158] Through the integration of the above-mentioned CNN backbone and its extension modules, a deep learning model is constructed that can fully perceive the characteristics of cardiac conduction signals.

[0159] Step S70: training a cardiac conduction health assessment model:

[0160] The purpose of this step is to train the constructed cardiac conduction health assessment model using the sample data collected in step S10. The specific steps are as follows:

[0161] 1. The cardiac conduction signal samples collected in step S10 and their corresponding expert health assessment data are divided into a training set and a validation set.

[0162] 2. Input the training set into the cardiac conduction health assessment model constructed in step S60 for end-to-end training. The training loss function may be mean square error (MSE) or cross entropy.

[0163] 3. Monitor the training effect of the model on the validation set. When the validation indicators are stable and optimal, save the trained model parameters.

[0164] 4. Through repeated iterative training, we finally get a deep learning model that can accurately predict the health of cardiac conduction.

[0165] Step S80: Output the cardiac conduction health assessment result:

[0166] The purpose of this step is to use the trained cardiac conduction health assessment model to predict and evaluate the newly input cardiac conduction signal. The specific steps are as follows:

[0167] 1. Obtain a sample of cardiac conduction signals to be evaluated.

[0168] 2. Input the sample into the heart conduction health assessment model trained in step S70.

[0169] 3. The model will automatically extract the standard features and non-standard features of the sample, and give the cardiac conduction health assessment result of the sample based on the learned health assessment rules.

[0170] 4. Output the evaluation results in the form of numerical values ​​or grades to provide a basis for doctors' diagnosis and treatment decisions.

[0171] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the above-mentioned method for assessing cardiac conduction health.

[0172] A third aspect of the present invention provides a cardiac conduction health assessment system, which includes the above-mentioned computer-readable storage medium.

[0173] Specifically, the principle of the present invention is:

[0174] 1. Multimodal physiological signal acquisition The present invention first collects a variety of cardiac conduction-related physiological signals including electrocardiogram, phonocardiogram, cardiac ultrasound and cardiac magnetic resonance. These signals can fully reflect the structural, functional, physiological and pathological characteristics of the cardiac conduction system. At the same time, expert evaluation data on the cardiac conduction health status of each subject is also collected to provide a benchmark standard for subsequent model training and performance evaluation.

[0175] 2. Separation of standard signals from non-standard signals In order to better analyze the health status characteristics contained in the cardiac conduction signal, the present invention uses the empirical mode decomposition (EMD) algorithm to decompose the collected original signal. EMD is an adaptive nonlinear signal decomposition method that can decompose complex signals into a series of intrinsic mode functions (IMFs) and a residual signal. The present invention classifies the part of the IMF that is most similar to the standard health state as a standard signal matrix, and classifies the dissimilar part as a non-standard signal matrix. This separation helps to better extract the characteristics of the healthy and abnormal states.

[0176] 3. Feature extraction of standard signals and non-standard signals For the two types of signals separated from the standard signal matrix and the non-standard signal matrix, the present invention uses wavelet transform and Fourier transform to extract features respectively. Wavelet transform can effectively capture the time-frequency characteristics of standard signals, reflecting the regularity and stability of cardiac conduction in a healthy state. Fourier transform can extract the spectral characteristics of non-standard signals, reflecting the frequency domain anomalies of signals under abnormal conditions. Through the extraction of such multi-dimensional features, the health status of the cardiac conduction system can be fully characterized.

[0177] 4. Construction and training of deep learning model Based on the above-extracted standard features and non-standard features, the present invention constructs a deep learning evaluation model that integrates modules such as convolutional neural network, attention mechanism, and residual connection. The model can effectively extract the spatiotemporal features of the signal, and adaptively highlight the key features through the attention mechanism, while using residual connection to enhance the gradient flow and feature learning capabilities. In addition, the present invention also uses spatial pyramid pooling and multi-scale feature fusion technology to further enhance the model's perception of features of different scales.

[0178] Through end-to-end training on a large amount of sample data, the deep learning model has learned the rules for extracting cardiac conduction health assessment features from standard and non-standard signals. During the training process, the model automatically optimizes internal parameters to make its predicted output as close as possible to the expert evaluation criteria. After training, the model can be used to assess the health of newly input cardiac conduction signals.

[0179] In order to better understand and implement the present invention, a specific embodiment of the present invention is provided below: the cardiology department of a large hospital is used as the first application scenario of the method of the present invention. The hospital has a complete heart health monitoring system, which regularly collects and analyzes various heart physiological signals of people with different health conditions. This embodiment is described based on the actual application case of the hospital.

[0180] The hospital selected a total of 120 subjects of different ages, genders and health conditions to participate in this experiment. Among them, 60 healthy people (without heart disease) and 60 patients with confirmed heart disease were included. A variety of cardiac conduction-related physiological signals such as 12-lead electrocardiogram, phonocardiogram, transthoracic ultrasound and cardiac magnetic resonance imaging were collected from each subject. At the same time, senior cardiologists of the hospital conducted professional assessments of the cardiac conduction health status of each subject based on the clinical diagnosis results and recorded the assessment level.

[0181] The basic information of the subjects is shown in Table 1. As can be seen, this experiment covers various cardiac conduction function states from normal to severely abnormal.

[0182] Table 1 Basic information statistics of subjects

[0183] Health status Number of subjects Average age Male to female ratio normal 60 45.2 1.2:1 Mild abnormality 28 52.6 1.4:1 Moderate abnormality 18 59.1 1.8:1 Severe abnormality 14 63.4 2.3:1

[0184] Figure 2 This is the age distribution of subjects in different health states. This chart uses a probability density function to show the age distribution of subjects in four different health states (normal, mildly abnormal, moderately abnormal, and severely abnormal). Each state is represented by a different color, and the mean age and standard deviation are shown. This chart intuitively shows the relationship between health status and age. It can be seen that with age, the proportion of abnormal states increases.

[0185] During the signal collection process, the hospital was equipped with professional physiological signal collection equipment. For electrocardiogram, a 12-lead Holter monitor was used to continuously record the subjects' electrocardiogram activities; for phonocardiogram, a high-sensitivity phonocardiogram sensor was used to collect acoustic signals generated by the opening and closing of each heart valve; for cardiac ultrasound, a color Doppler ultrasound instrument was used to dynamically observe the structure and function of each chamber of the subject's heart; for cardiac magnetic resonance, a 1.5T magnetic resonance imaging device was used to scan the subject's heart to obtain data, as shown in Table 2.

[0186] Table 2 Parameters of physiological signal acquisition equipment

[0187] Signal Type Collection equipment Sampling rate Resolution Electrocardiogram Holter Monitor 1000Hz 16bit Phonocardiogram Heart sound sensor 2000Hz 24bit Cardiac ultrasound Color Doppler Ultrasound 50Hz 12bit Cardiac MRI 1.5T MRI - -

[0188] After completing the sample data collection, the research team first preprocessed the original signal. This includes eliminating power supply interference, filtering out physiological noise such as breathing, and eliminating quantization noise. The preprocessed signal is clearer, which is conducive to subsequent feature extraction and model training.

[0189] By comparison, it can be seen that the noise of the ECG signal after preprocessing is significantly suppressed, and the waveform features are clearer, which lays a good foundation for subsequent feature extraction and health status analysis.

[0190] Next, the research team applied the empirical mode decomposition (EMD) algorithm to decompose all the collected physiological signal sample data. EMD is an adaptive nonlinear signal decomposition method that can decompose complex original signals into a series of intrinsic mode functions (IMFs) and a residual signal. Through this decomposition, the standard signal components and non-standard signal components contained in the signal can be separated.

[0191] Figure 3 The process and results of empirical mode decomposition (EMD) are shown. The top shows the original ECG signal, and below are three intrinsic mode functions (IMFs) and a residual signal. Each IMF represents a frequency component in the signal, arranged from high frequency to low frequency. This chart helps to understand how EMD decomposes complex ECG signals into simpler components for subsequent analysis. From the EMD decomposition results, it can be seen that compared with healthy people, the ECG signals of heart disease patients have more non-standard components. This difference may be due to abnormal cardiac conduction function, which in turn causes the signal to present more irregular characteristics. Therefore, further analysis of these standard signal and non-standard signal components will help to better assess the health of cardiac conduction.

[0192] Next, the research team calculated the features extracted from the standard signal matrix and the non-standard signal matrix respectively. For the standard signal, the wavelet transform was used to extract the time-frequency features; for the non-standard signal, the Fourier transform was used to extract the spectrum features.

[0193] Wavelet transform is a signal processing method that can simultaneously reflect the characteristics of signals in the time domain and frequency domain. Its mathematical expression is: Where, W f (a, b) is the wavelet transform coefficient, f(t) is the input signal, ψ(t) is the wavelet function, a is the scale parameter, and b is the translation parameter. Through wavelet transform, the frequency characteristics of the standard signal at different scales and times can be obtained.

[0194] Figure 4 The wavelet time-frequency diagram of the ECG of healthy subjects and heart patients is compared. Through this comparison, it can be clearly seen that the ECG signal of healthy subjects should show regular and clear QRS complexes; the wavelet time-frequency diagram of healthy subjects should show regular and periodic energy distribution patterns. The ECG signal of heart patients may show irregular RR intervals and additional fluctuations. The wavelet time-frequency diagram of heart patients may show more irregular and scattered energy distribution, reflecting unstable heart rate and possible conduction abnormalities.

[0195] For non-standard signals, the research team used Fourier transform to extract spectral features. Fourier transform can convert time domain signals into frequency domain, reflecting the energy distribution of the signal in different frequency components. Its mathematical expression is: Where F(ω) is the frequency spectrum of the signal f(t), ω is the angular frequency, and j is the imaginary unit.

[0196] The structure of the deep learning model from left to right is the input layer, convolution layer, pooling layer, spatial pyramid pooling (SPP) layer, attention mechanism module, residual connection module, fully connected layer and output layer. The main body of the model adopts the convolutional neural network structure, which can effectively extract the spatiotemporal characteristics of the signal. In order to enhance the perception ability of the model, the following key modules are integrated into the basic CNN network:

[0197] 1. Spatial Pyramid Pooling (SPP) layer: It replaces the original pooling layer and can adaptively extract multi-scale features and enhance the model's ability to learn features of different scales.

[0198] 2. Attention mechanism module: By adaptively assigning weights to different features, the expression of key features is highlighted and the discriminability of the model is improved.

[0199] 3. Residual connection module: Enhances the flow of gradients, which helps the model better learn deep-level features related to cardiac conduction.

[0200] Through end-to-end training on a large amount of sample data, the model can learn the rules for extracting cardiac conduction health assessment features from standard and non-standard signals. During the training process, the model automatically optimizes internal parameters to make its predicted output as close as possible to the expert evaluation criteria. After training, the model can be used to assess the health status of newly input cardiac conduction signals.

[0201] In practical applications, hospitals can use this deep learning model to quickly evaluate the collected cardiac conduction signals. The operation process is as follows:

[0202] 1. Collect the subject's electrocardiogram, phonocardiogram, cardiac ultrasound, cardiac magnetic resonance and other physiological signals.

[0203] 2. Input the collected signal into the trained deep learning model for prediction.

[0204] 3. The model will output the cardiac conduction health assessment results of the subject, such as normal, mildly abnormal, moderately abnormal or severely abnormal.

[0205] 4. The doctor conducts further analysis and diagnosis based on the evaluation results and clinical diagnostic data.

[0206] Table 3 lists the results of the cross-validation test conducted by the hospital on a sample of 120 subjects. It can be seen that the deep learning model has achieved high accuracy and specificity in health status assessment, providing strong support for the accurate monitoring of heart health.

[0207] Table 3 Performance indicators of cardiac conduction health evaluation model

[0208] index value Accuracy 91.7％ Sensitivity 89.3％ Specificity 93.5％ ROC-AUC 0.946

[0209] Figure 5 It consists of two parts. On the left is a bar graph showing the accuracy, sensitivity, and specificity of the model. On the right is the ROC (Receiver Operating Characteristic) curve, which shows the performance of the model at different thresholds and gives the AUC (Area Under the Curve) value. This comprehensive chart comprehensively displays the performance indicators of the model and helps to evaluate the effectiveness of the model.

[0210] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for evaluating cardiac conduction health, characterized in that: The following steps are involved: S10, obtaining a plurality of cardiac conduction signals of a plurality of test persons and obtaining expert evaluation data of the cardiac conduction health of the test persons; S20, performing decomposition calculation on each signal respectively to obtain a standard signal matrix and a non-standard signal matrix; S30, according to a preset splitting threshold, determining a signal in the standard signal matrix whose similarity to the corresponding standard signal is higher than the splitting threshold; S40, using a wavelet transform method to calculate the characteristics of the signals in the standard signal matrix to obtain standard characteristics; S50, using a Fourier transform method to calculate the characteristics of the signal in the non-standard signal matrix to obtain non-standard characteristics; S60, constructing a cardiac conduction health assessment model based on the standard features and the non-standard features; S70, training the heart conduction health assessment model to obtain a trained heart conduction health assessment model; S80, obtaining a cardiac conduction signal to be evaluated, inputting the cardiac conduction signal to be evaluated into the trained cardiac conduction health assessment model, obtaining a cardiac conduction health assessment result and outputting it.

2. A cardiac conduction health assessment method according to claim 1, characterized in that: The multiple signals conducted by the heart include electrocardiogram signals, heart sound signals, cardiac ultrasound signals and cardiac magnetic resonance signals.

3. A cardiac conduction health assessment method according to claim 2, characterized in that: The test subjects included people with different health conditions as well as patients with existing heart conditions.

4. A cardiac conduction health assessment method according to claim 3, characterized in that: The cardiac conduction health assessment model specifically adopts a convolutional neural network structure, replaces the pooling layer therein with a spatial pyramid pooling layer, and adds an attention mechanism module, a residual connection module and a multi-scale feature fusion module to obtain the cardiac conduction health assessment model.

5. A cardiac conduction health assessment method according to claim 4, characterized in that: The decomposition calculation specifically uses the empirical mode decomposition method to decompose the signal.

6. A cardiac conduction health assessment method according to claim 5, characterized in that: The similarity adopts cosine similarity.

7. A cardiac conduction health assessment method according to claim 6, characterized in that: The step of obtaining the splitting threshold comprises: Step 1: Collect a large number of cardiac conduction signal samples from healthy people and patients with known heart diseases; Step 2: pre-process the sample; Step 3, calculating the similarity between the sample signal and the standard signal; Step 4: Use the support vector machine algorithm to classify the similarities; Step 5: Determine the optimal split threshold through cross-validation and ROC curve analysis.

8. A cardiac conduction health assessment method according to claim 7, characterized in that: The formula for preprocessing the sample is: In the formula, x norm is the standardized signal; x is the original signal; μ is the mean of the original signal; σ is the standard deviation of the original signal.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute a cardiac conduction health assessment method as described in any one of claims 1-8.

10. A cardiac conduction health assessment system, characterized in that: A computer-readable storage medium comprising the computer-readable storage medium of claim 9.

Images