Electrocardiosignal quality evaluation method based on unsupervised cascade adaptive network

Abstract

The invention relates to the technical field of electrocardiosignal quality evaluation, and discloses an electrocardiosignal quality evaluation method based on an unsupervised cascade adaptive network, which comprises the following steps: S1, preprocessing an obtained original electrocardiosignal: removing power frequency interference through adaptive notch filtering, removing high-frequency myoelectricity noise through soft threshold wavelet filtering, and obtaining a pre-processed electrocardiosignal; keeping a potential pathological characteristic waveform in a filtering process to obtain a preprocessed electrocardiosignal; s2, unsupervised multi-modal feature extraction: an unsupervised feature extraction network is constructed based on an auto-encoder; a normal electrocardiosignal distribution rule is automatically learned through an auto-encoder, so that the dependence on labeling resources of professional doctors is reduced; meanwhile, by means of a two-stage self-adaptive evaluation network, significant noise is filtered first, then pathological signals and noise are accurately distinguished, it is effectively avoided that abnormal waveforms caused by diseases such as myocardial infarction and atrial fibrillation are misjudged as noise, the retention rate of the pathological signals is remarkably increased, and more complete effective data support is provided for clinical diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electrocardiosignal quality evaluation, in particular to an electrocardiosignal quality evaluation method based on an unsupervised cascade adaptive network. BACKGROUND

[0002] Electrocardiosignal quality evaluation is a key link in cardiovascular disease diagnosis and remote monitoring. Its core is to screen out effective signals without interference and accurately reflecting the heart function state through technical means, to provide reliable data basis for subsequent diagnosis and treatment work such as electrocardiogram analysis and arrhythmia detection, and is widely used in hospital clinical diagnosis, family health monitoring and wearable medical devices and other scenes.

[0003] In the current industry, electrocardiosignal quality evaluation mostly adopts a technical path combining supervised learning with single feature extraction, that is, a large number of "normal / noise" electrocardiosignal samples are manually labeled to train a model, time-frequency domain features of the signal are extracted by methods such as wavelet transform and Fourier transform, and quality classification is completed by models such as support vector machine and single convolutional neural network. Although some methods introduce unsupervised learning, they mostly adopt a single-stage evaluation mode and directly make binary determination of "qualified / unqualified" on the signal.

[0004] However, the traditional method is easy to misidentify the signal waveform abnormality caused by heart disease as noise interference, resulting in the loss of important clinical information. For example, some rule-based methods detect noise by setting a fixed threshold. When a patient with heart disease has an abnormal electrocardiosignal, these abnormalities are likely to be incorrectly determined as low-quality signals and excluded. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides an electrocardiosignal quality evaluation method based on an unsupervised cascade adaptive network, which solves the problem of misjudging electrocardiosignal disease waveform as noise.

[0006] To achieve the above purpose, the present application is implemented by the following technical scheme: an electrocardiosignal quality evaluation method based on an unsupervised cascade adaptive network, comprising the following steps: S1, preprocessing the obtained original electrocardiosignal: removing power frequency interference by adaptive notch filtering, removing high-frequency electromyographic noise by soft threshold wavelet filtering, and retaining potential pathological feature waveforms in the filtering process to obtain a preprocessed electrocardiosignal; S2, unsupervised multi-modal feature extraction: an unsupervised feature extraction network based on a self-encoder is constructed to extract time domain features, frequency domain features and nonlinear features from the preprocessed electrocardiosignal of S1, and fuse the three types of features to generate a comprehensive feature vector; S3, cascade adaptive evaluation: a two-stage adaptive evaluation network is constructed to process the comprehensive feature vector, specifically including: First level noise filtering: using improved K-means algorithm to cluster and filter the significant noise features in the comprehensive feature vector, eliminate obvious noise signals, and output candidate signals; Second level pathological-noise distinction: introducing an adaptive attention mechanism, using a pathological feature confidence calculation algorithm to preliminarily classify the candidate signals output by the first level, outputting a preliminary quality assessment result, which includes three categories: suspected pathological signals, noise, and ambiguous interval signals; at the same time, the ambiguous interval signals are subdivided by a secondary classification algorithm, and a subdivision result is output, which includes two categories: weak pathological signals and pseudo-pathological noise; S4, output quality level: combining the preliminary quality assessment result and the subdivision result, and a preset clinical pathological feature library, output a complete electrocardiogram signal quality assessment result, which includes four categories: suspected pathological signals, weak pathological signals, pseudo-pathological noise, and noise.

[0007] Preferably, the threshold calculation method of the soft threshold wavelet filtering in the S1 central electrical signal preprocessing module is: estimating the noise standard deviation based on the high frequency coefficient after wavelet decomposition, combining the sampling point number of the original electrocardiogram signal, and determining the threshold value through a logarithmic relationship to ensure that the high frequency electromyographic noise with a frequency greater than 100 Hz is removed while the low frequency pathological features are not lost.

[0008] Preferably, the extraction method of time domain features in the S2 unsupervised multi-modal feature extraction module is: using the sliding window method to intercept the waveform fragments of the preprocessed electrocardiogram signal, calculating the QRS wave amplitude, RR interval standard deviation, and ST segment offset of each fragment, a total of 3 time domain parameters; the extraction method of frequency domain features is: performing fast Fourier transform on the preprocessed electrocardiogram signal in S1, calculating the power spectrum entropy, and the low frequency and high frequency energy ratio of heart rate variability, a total of 2 frequency domain parameters; the extraction method of nonlinear features is: calculating the fractal dimension using the box counting method, and calculating the sample entropy using the sample entropy algorithm, a total of 2 nonlinear parameters; a total of 7 parameters are fused to generate a comprehensive feature vector with equal weight.

[0009] Preferably, the network structure parameters of the autoencoder in the S2 unsupervised multi-modal feature extraction module are: the coding layer dimension is set to 128, the decoding layer dimension is consistent with the dimension of the preprocessed electrocardiogram signal, the mean square error is used as the loss function, and the distribution rule of normal electrocardiogram signal is learned by minimizing the loss function, providing a reference for the subsequent pathological feature confidence calculation algorithm.

[0010] Preferably, the pathological feature confidence calculation algorithm of the second level pathological-noise distinction sub-network in the cascaded adaptive evaluation module is used to preliminarily classify the candidate signals, and the specific calculation formula includes:

[0011] In the formula: : pathological feature confidence, the value range is [0, 1], as a preliminary classification basis; when ≥ 0.6, it is preliminarily determined as a suspected pathological signal, when <0.3, it is preliminarily determined as noise, and when 0.3≤ <0.6, it is marked as a fuzzy interval signal; : the weight of the i-th pathological correlation feature, a total of 3 core pathological correlation features, which are ST segment offset, fractal dimension, and sample entropy, and the corresponding weights are 0.4, 0.3, and 0.3, respectively. The weight is obtained by training 500 myocardial infarction samples and 300 atrial fibrillation samples in a preset clinical pathological feature library; : the normalized value of the i-th pathological correlation feature, the value range is [0, 1], when the ST segment offset is > 0.1 mV, the normalized value is 1, when the fractal dimension is < 1.0, the normalized value is 1, and when the sample entropy is < 0.5, the normalized value is 1; : reconstruction error correction coefficient, the value is 0.2, used to balance the influence of the signal reconstruction result of unsupervised learning and the pathological feature weight; : reconstruction error of the autoencoder, the value range is [0, 0.1], the pathological waveform deviates from the normal signal distribution, resulting in a reconstruction error higher than that of the normal signal and lower than that of the random noise, wherein the normal signal reconstruction error is < 0.02, the pathological signal reconstruction error is 0.02~0.05, and the noise reconstruction error is > 0.05. : the preliminary classification threshold of the pathological feature confidence is determined by the receiver operating characteristic curve optimization: taking the pathological waveforms and noise marked in the preset clinical pathological feature library as the gold standard, adjusting the threshold to make the overall accuracy of the preliminary classification ≥ 90%, and reserving a subdivision proportion of 10%~15% of the total signal for the fuzzy interval signal.

[0012] Preferably, the preliminary classification threshold of the pathological feature confidence is determined by the receiver operating characteristic curve optimization: taking the pathological waveforms and noise marked in the preset clinical pathological feature library as the gold standard, adjusting the threshold to make the overall accuracy of the preliminary classification ≥ 90%, and reserving a subdivision proportion of 10%~15% of the total signal for the fuzzy interval signal. Preferably, the construction method of the preset clinical pathological feature library is as follows: 1000 clinically diagnosed pathological electrocardiogram samples and 500 normal electrocardiogram samples are collected, wherein the pathological electrocardiogram samples include 500 myocardial infarction samples, 300 atrial fibrillation samples, and 200 ventricular premature beat samples; waveform feature parameters are marked for each sample, including QRS wave width, ST segment elevation amplitude, and RR interval regularity; three types of typical pathological waveform templates are generated by K-means clustering, and the parameters of the three types of typical pathological waveform templates are:

[0013] ​The first type of template corresponds to myocardial infarction, and the parameter is that the ST segment elevation amplitude is greater than or equal to 0.1 mV in the chest lead or greater than or equal to 0.2 mV in the limb lead, and the QRS wave is QS type; The second type of template corresponds to atrial fibrillation, and the parameter is that the RR interval is absolutely irregular, the f wave frequency is 350-600 times per minute, and the f wave amplitude is less than 0.1 mV; The third type of template corresponds to ventricular premature beat, and the parameter is that the QRS wave width is greater than 120 ms, and the compensation interval is complete; The preset clinical pathological feature library is verified by clinical verification of a third-grade hospital, and the matching accuracy of the template and the real pathological waveform is greater than or equal to 95%.

[0014] Preferably, the secondary classification algorithm of the second pathological-noise distinguishing module network in the cascade adaptive evaluation module is used to subdivide the fuzzy interval signal, and the subdivision result output includes two types of weak pathological signal and pseudo-pathological noise, and the specific calculation formula includes:

[0015] In the formula: The secondary classification confidence is in the range of [0, 1], and is used as the subdivision basis; when ≥0.5, it is determined as a weak pathological signal, and when <0.5, it is determined as a pseudo-pathological noise; The dispersion of pathological correlation characteristics is in the range of [0, 1], and the calculation method is the ratio of the standard deviation of three core pathological correlation characteristics (ST segment offset, fractal dimension, and sample entropy) to the maximum value of each feature; The matching residual error of the signal and the preset clinical pathological feature library template is in the range of [0, 1], and the calculation method is to normalize the Euclidean distance between the comprehensive feature vector of the fuzzy interval signal and the most similar template feature vector in the preset clinical pathological feature library to the interval [0, 1]; The secondary classification balance coefficient is 0.4, which is used to balance the weights of the feature dispersion, the matching residual error and the pathological feature confidence C; The pathological feature confidence is the preliminary classification result output by the pathological feature confidence calculation algorithm, and the value range is 0.3≤ <0.6.

[0016] Preferably, the screening rule of the most similar template in the secondary classification algorithm is that the Euclidean distance of the comprehensive feature vector F_candidate of the fuzzy interval signal and the feature vectors of the three types of templates in the preset clinical pathological feature library is calculated, and the template with the smallest distance is selected as F_template; if the smallest Euclidean distance is greater than a preset threshold, which is 1.2 times of the average Euclidean distance of all template feature vectors in the preset clinical pathological feature library, then e_match is forced to be 0.8.

[0017] Preferably, the result fusion rule of the quality level output module is that the suspected pathological signals and noises determined by the pathological feature confidence calculation algorithm are retained, the weak pathological signals and pseudo-pathological noises output by the secondary classification algorithm are added, and four complete classifications are formed; wherein the weak pathological signals need to be associated with the corresponding template type in the preset clinical pathological feature library to provide reference basis for clinical diagnosis.

[0018] The application provides an electrocardiosignal quality evaluation method based on an unsupervised cascade adaptive network. 1、The application automatically learns the normal electrocardiosignal distribution rule through the autoencoder, reduces the dependence on professional doctor annotation resources, and simultaneously filters significant noises through the two-level adaptive evaluation network, accurately distinguishes pathological signals from noises, effectively avoids misjudging abnormal waveforms caused by myocardial infarction, atrial fibrillation and other diseases as noises, and significantly improves the pathological signal retention rate, thereby providing more complete and effective data support for clinical diagnosis.

[0019] 2、The application comprehensively captures the time domain, frequency domain and nonlinear features of the electrocardiosignal through multi-modal feature fusion and double algorithm cooperation, supplements the feature dispersion degree and template matching dimension judgment for the fuzzy interval signal, improves the evaluation accuracy, and the preset clinical pathological feature library is clinically verified, is suitable for common cardiovascular disease signal recognition, and takes into account the real-time performance of wearable devices and the practicability of clinical diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The flowchart of the application. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the specification of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0022] Embodiment: Please refer to the drawings in the specification of the application Figure 1The embodiment of the application provides a kind of electrocardiosignal quality evaluation method based on unsupervised cascade adaptive network, comprising the following steps: S1, the original electrocardiosignal obtained is preprocessed: remove power frequency interference by adaptive notch filter, remove high-frequency electromyographic noise by soft threshold wavelet filter, and retain potential pathological characteristic waveform in the filtering process, to obtain preprocessed electrocardiosignal; Adopt 250Hz sampling frequency, 16-bit precision dynamic electrocardiogram recorder to collect 30 minutes Ⅱ lead signal (EDF format);Adaptive notch filter is IIR filter, real-time matching 50 / 60Hz power frequency, bandwidth 2Hz, interference amplitude is reduced to less than 5% of original signal;Soft threshold wavelet filter selects db4 wavelet base to do 5 layers decomposition, only processes 1-3 layers high frequency coefficient (>100Hz), 4-5 layers low frequency coefficient (including QRS wave, ST segment) is retained;Confirm the pathology characteristics by comparing the QRS waveform (peak-valley position deviation ≤5ms), and output the preprocessed signal.

[0023] S2, unsupervised multi-modal feature extraction: based on the unsupervised feature extraction network constructed by self-encoder, the electrocardiosignal preprocessed by S1 is extracted respectively in time domain, frequency domain and nonlinear domain, and the three kinds of features are fused to generate a comprehensive feature vector; The input layer dimension of the self-encoder is adapted to the length of the preprocessed electrocardiosignal. The dimension of the encoding layer can be set to 128, and the ReLU activation function is used;The decoding layer uses Sigmoid activation function, and its dimension is consistent with the input layer. The mean square error is used as the loss function, and a certain number of healthy person electrocardiosignal samples (such as 100) are used to train the network. The network is trained for multiple iterations (such as 500 times), the learning rate can be set to 0.001, until the loss function converges to a lower value (such as less than 0.01), so as to learn the distribution rule of normal electrocardiosignal, and the signal segment is intercepted by sliding window method to extract features. The length of the sliding window can be 2 seconds, and the step can be 1 second. 7 feature parameters (describe feature types) in each window are calculated. After normalizing all parameters, they are fused with equal weight to generate the corresponding comprehensive feature vector.

[0024] S3, cascade adaptive evaluation: construct two-stage adaptive evaluation network to process the comprehensive feature vector, specifically including: First level noise filtering: use improved K-means algorithm to cluster and filter the significant noise features in the comprehensive feature vector, eliminate obvious noise signals, and output candidate signals; Second level pathological-noise discrimination: introduce an adaptive attention mechanism, and perform preliminary classification on the candidate signals output by the first level through a pathological feature confidence calculation algorithm to output a preliminary quality assessment result, the preliminary quality assessment result including three categories of suspected pathological signals, noise and ambiguous interval signals; meanwhile, the ambiguous interval signals are subdivided through a secondary classification algorithm to output a subdivision result, the subdivision result including two categories of weak pathological signals and pseudo-pathological noise; Improved K-means with K-means++ setting 2 cluster centers (significant noise / effective signal), iterating 50 times; calculating the Euclidean distance of the vector and the center, and removing the segments deviating to the noise center by more than 50%; the length of the candidate signal needs to be greater than 60% of the original signal, otherwise it is determined as a whole low quality.

[0025] S4, output quality level: combining the preliminary quality assessment result and the subdivision result, and a preset clinical pathological feature library, output a complete electrocardio signal quality assessment result, the complete electrocardio signal quality assessment result including four categories of suspected pathological signals, weak pathological signals, pseudo-pathological noise and noise.

[0026] Calling a feature library containing three types of templates of myocardial infarction, atrial fibrillation and premature beat, calculating the Euclidean distance of the signal and the template to obtain the optimal matching; outputting a table containing: overall level (suspected pathological proportion greater than 30% as "containing pathological information" and other four categories), time distribution of each signal (accurate to seconds), matching template and matching degree; the result is synchronized to a clinical diagnosis system.

[0027] In the implementation of the present application, the threshold calculation method of the soft threshold wavelet filtering in the S1 center electrical signal preprocessing module is: based on the estimation of the noise standard deviation of the high frequency coefficient after wavelet decomposition, combining the sampling point number of the original electrocardio signal, determining the threshold value through the logarithmic relationship, to ensure that the high frequency myoelectric noise with a frequency greater than 100 Hz is removed while the low frequency pathological features are not lost.

[0028] The specific steps are as follows: 1. Use db4 wavelet base to decompose the original signal for 5 layers, and obtain 5 groups of high frequency coefficients and 1 group of low frequency coefficients after decomposition, and extract the first layer of high frequency coefficients (this layer of coefficients mainly contains high frequency noise components in the signal); 2. Estimate the fluctuation degree of the noise (i.e. the noise standard deviation) by counting the distribution of the first layer of high frequency coefficients; 3. Calculate the filtering threshold value by using the logarithmic relationship in combination with the total sampling point number of the original signal (sampling at 250 Hz for 30 minutes, total point number is 7500); 4. Process the first to third layer of high frequency coefficients (all high frequency components greater than 100 Hz) with the threshold value, and retain the fourth to fifth layer of low frequency coefficients (containing low frequency pathological features such as QRS wave and ST segment), to achieve the balance between high frequency noise filtering and pathological feature retention.

[0029] In the implementation of the present application, the extraction method of time domain features in the unsupervised multi-modal feature extraction module in S2 is: the waveform fragments of the preprocessed electrocardiogram signal are intercepted by the sliding window method, the QRS wave amplitude, the RR interval standard deviation, and the ST segment offset of each fragment are calculated, and a total of 3 time domain parameters are obtained; the extraction method of frequency domain features is: the preprocessed electrocardiogram signal in S1 is subjected to fast Fourier transform, the power spectrum entropy, the low-frequency and high-frequency energy ratio of heart rate variability are calculated, and a total of 2 frequency domain parameters are obtained; the extraction method of nonlinear features is: the fractal dimension is calculated by using the box counting method, and the sample entropy is calculated by using the sample entropy algorithm, and a total of 2 nonlinear parameters are obtained; a total of 7 parameters are fused to generate a comprehensive feature vector with equal weights.

[0030] The sliding window is set to 2 seconds, and each time it is slid by 1 second to ensure that the entire signal is covered and the fragments are not repeated; the QRS wave amplitude is the difference between the R wave peak value and the S wave valley value in each window, the RR interval standard deviation is the dispersion degree of all adjacent heartbeat intervals in the window, and the ST segment offset is the potential at 80 ms after the J point compared with the PR segment reference potential; 1024 points are taken when performing fast Fourier transform, the power spectrum entropy reflects the uniformity of the energy distribution of the signal in the 0.5-40 Hz frequency band, and the low-frequency (0.04-0.15 Hz) and high-frequency (0.15-0.4 Hz) energy ratios of heart rate variability reflect the autonomic nervous function; the box counting method uses a box length of 0.01-0.1 mV to cover the waveform, and the sample entropy is calculated by measuring the regularity of signal fluctuations; the 7 parameters are first adjusted to between 0 and 1, and then added with a weight of one-seventh each to form a comprehensive feature vector.

[0031] In the implementation of the present application, the network structure parameters of the autoencoder in the unsupervised multi-modal feature extraction module in S2 are: the coding layer dimension is set to 128, the decoding layer dimension is consistent with the dimension of the preprocessed electrocardiogram signal, the mean square error is used as the loss function, and the distribution rule of the normal electrocardiogram signal is learned by minimizing the loss function to provide a reference for the subsequent pathological feature confidence calculation algorithm.

[0032] The number of neurons in the input layer of the autoencoder is consistent with the length of the preprocessed signal (7500 neurons corresponding to 30 minutes of 250 Hz sampling), the coding layer has 128 neurons, the ReLU activation function is used to enhance the feature extraction capability, and the number of neurons in the decoding layer is the same as that in the input layer, and the Sigmoid activation function is used to make the output signal range match the original signal; during training, the electrocardiogram data of 100 people without cardiovascular diseases is used, the network parameters are adjusted each time, and when the mean square error (reflecting the difference between the network output and the input signal) is reduced to below 0.01, the training is stopped, at this time the network can accurately reproduce the waveform rule of the normal electrocardiogram signal, and provide a reliable normal signal reference for subsequent differentiation of "normal signal, pathological signal, noise".

[0033] In this invention, the pathological feature confidence calculation algorithm of the second-level pathological-noise region molecular network in the cascaded adaptive evaluation module is used to achieve preliminary classification of candidate signals. The specific calculation formula includes:

[0034] In the formula: : Confidence level of pathological features, with a value range of [0,1], used as a preliminary classification basis; when A value ≥0.6 is initially considered a suspected pathological signal. If the value is less than 0.3, it is initially determined to be noise; if 0.3 ≤ Signals with a value less than 0.6 are marked as fuzzy interval signals; The weight of the i-th pathological association feature includes three core pathological association features: ST segment offset, fractal dimension, and sample entropy, with corresponding weights of 0.4, 0.3, and 0.3, respectively. The weights are obtained by training with 500 myocardial infarction samples and 300 atrial fibrillation samples from a pre-set clinical pathological feature library. : No. The normalized value of each pathological correlation feature ranges from [0,1]. When the ST segment offset is greater than 0.1mV, the normalized value is 1. When the fractal dimension is less than 1.0, the normalized value is 1. When the sample entropy is less than 0.5, the normalized value is 1. : Reconstruction error correction coefficient, with a value of 0.2, is used to balance the influence of signal reconstruction results from unsupervised learning and the weights of pathological features; The reconstruction error of the autoencoder ranges from [0, 0.1]. The reconstruction error of the pathological waveform is higher than that of the normal signal but lower than that of the random noise due to deviation from the normal signal distribution. The reconstruction error of the normal signal is <0.02, the reconstruction error of the pathological signal is 0.02~0.05, and the reconstruction error of the noise is >0.05.

[0035] The calculation first takes the normalized values ​​of the three pathological features of the candidate signal and calculates the weighted sum according to the set weights. Then, the reconstruction effect of the autoencoder on the signal is observed (the smaller the reconstruction error, the closer the signal is to normal). After adjusting with correction coefficients, it is combined with the weighted sum to obtain a confidence level between 0 and 1. For example, a signal has an ST segment offset of 0.12mV, a fractal dimension of 0.9, and a sample entropy of 0.4 (all three normalized values ​​are 1), a reconstruction error of 0.03, and a weighted sum of 0.4 + 0.3 + 0.3 = 1. Combined with the correction coefficient, the confidence level is calculated to be about 0.86, which is judged as a suspected pathological signal. The weights of the three features are determined through repeated debugging of 800 pathological samples to ensure the sensitivity of identification of pathological features such as myocardial infarction and atrial fibrillation.

[0036] In the implementation of the present application, the confidence of the pathological feature The preliminary classification threshold is determined by optimizing the receiver operating characteristic curve: taking the labeled pathological waveforms and noise in the preset clinical pathological feature library as the gold standard, adjusting the threshold to make the overall accuracy of the preliminary classification ≥ 90%, and reserving a subdivision proportion of 10%-15% of the total signal for the fuzzy interval signal.

[0037] 500 cases of clearly labeled pathological waveforms (including myocardial infarction, atrial fibrillation, and premature ventricular contraction) and 300 cases of noise signals in the feature library are selected as reference standards, and the receiver operating characteristic curve is drawn. By adjusting the confidence threshold, the "pathological signal recognition rate" and "noise recognition rate" are both at a high level; finally, when the confidence is ≥ 0.6, the pathological signal recognition rate is over 95%; when the confidence is < 0.3, the noise recognition rate is over 92%, the overall classification accuracy is 91%, and the proportion of fuzzy signals between 0.3 and 0.6 is about 12%, which neither increases the subsequent processing pressure due to too many fuzzy signals nor misses potential pathological information due to too few fuzzy signals.

[0038] In the implementation of the present application, the preset clinical pathological feature library is constructed as follows: 1000 cases of clinically diagnosed pathological electrocardiogram samples and 500 cases of normal electrocardiogram samples are collected, including 500 cases of myocardial infarction samples, 300 cases of atrial fibrillation samples, and 200 cases of premature ventricular contraction samples; waveform feature parameters are labeled for each sample, including QRS wave width, ST segment elevation amplitude, and RR interval regularity; three types of typical pathological waveform templates are generated by K-means clustering, and the parameters of the three types of typical pathological waveform templates are as follows: The first type of template corresponds to myocardial infarction, and its parameters are ST segment elevation amplitude ≥ 0.1 mV in chest leads or ≥ 0.2 mV in limb leads, and QRS wave is QS type; The second type of template corresponds to atrial fibrillation, and its parameters are absolute irregularity of RR interval, f wave frequency of 350-600 beats / min, and f wave amplitude < 0.1 mV; The third type of template corresponds to premature ventricular contraction, and its parameters are QRS wave width > 120 ms and complete compensatory pause; The preset clinical pathological feature library is verified by clinical trials in third-grade hospitals to ensure that the matching accuracy of the templates and the real pathological waveforms is ≥ 95%.

[0039] The samples are from the cardiology departments of three first-class hospitals, and are diagnosed and signed by two deputy chief physicians or above to avoid diagnostic errors; professional electrocardiogram analysis software is used for labeling, the QRS wave width is accurate to milliseconds, and the ST segment elevation amplitude is accurate to 0.01 mV; three cluster centers are set in K-means clustering, respectively corresponding to three types of pathology, and the clustering result is stable after 30 iterations, and the waveform similarity of the same sample is more than 96%; 200 pathological samples not involved in the database are used for clinical verification, and are compared with the three types of templates, and the average matching accuracy is 95.8%, among which the template matching accuracy of the heart block is the highest (97.2%), which can meet the needs of clinical recognition of typical pathological signals.

[0040] In the implementation of the application, the secondary classification algorithm of the second pathological-noise distinguishing network in the cascade adaptive evaluation module is used to subdivide the fuzzy interval signal, and the subdivision result output includes two types of weak pathological signals and pseudo-pathological noise, and the specific calculation formula includes:

[0041] In the formula: The secondary classification confidence is in the range of [0, 1], and is used as the subdivision basis; when C 2 >= 0.5, it is determined as a weak pathological signal, and when C 2 < 0.5, it is determined as a pseudo-pathological noise. <0.5. The dispersion of the pathological correlation feature is in the range of [0, 1], and the calculation method is the ratio of the standard deviation of three core pathological correlation features (ST segment offset, fractal dimension, and sample entropy) to the maximum value of each feature. The signal and the matching residual error of the preset clinical pathological feature library template are in the range of [0, 1], and the calculation method is the Euclidean distance between the comprehensive feature vector of the fuzzy interval signal and the most similar template feature vector in the preset clinical pathological feature library, and then normalized to the [0, 1] interval. The secondary classification balance coefficient is 0.4, which is used to balance the weights of the feature dispersion, the matching residual error and the pathological feature confidence C. The pathological feature confidence is the preliminary classification result output by the pathological feature confidence calculation algorithm, and the value range is 0.3 <= C 1 <= 0.6. <0.6.

[0042] ​The calculation is first to see the discrete degree of three pathological characteristics (the smaller the dispersion, the more stable the characteristics, and the more likely it is a real pathological signal), and then to see the similarity of the signal and the template of the feature library (the smaller the matching residual, the closer to the known pathological waveform), and to combine the two new indicators with the preliminary confidence by using the balance coefficient to obtain the secondary classification confidence; for example, a fuzzy signal characteristic has a dispersion of 0.05, a matching residual of 0.32, a preliminary confidence of 0.45, a balance coefficient of 0.4, and a final secondary confidence of about 0.52, which is determined as a weak pathological signal; the balance coefficient 0.4 is determined by 100 fuzzy signal debugging, so that the new indicator and the original confidence cooperate to improve the subdivision accuracy.

[0043] In the implementation of the application, the screening rule of the most similar template in the secondary classification algorithm is: calculate the Euclidean distance of the comprehensive feature vector F candidate of the fuzzy interval signal and the three types of template feature vectors in the preset clinical pathological feature library, and select the template with the smallest distance as F template; if the smallest Euclidean distance is greater than a preset threshold, which is 1.2 times the average Euclidean distance of all template feature vectors in the preset clinical pathological feature library, then e match is forced to be 0.8.

[0044] First, calculate the distance of the fuzzy signal vector and the three types of template vectors of myocardial infarction, atrial fibrillation and premature beat, and select the template with the smallest distance as the most similar template; for example, the distances of a signal and the three types of templates are 0.28, 0.55 and 0.42 respectively, the smallest distance 0.28 corresponds to the myocardial infarction template, which is the most similar template; first, calculate the average distance between the three types of templates in the feature library (for example, 0.25), and set 1.2 times (0.3) of the average distance as the threshold; if the distance of the signal and the most similar template is 0.28 (less than 0.3), calculate the matching residual according to the actual distance; if the distance is 0.35 (greater than 0.3), it means that the signal is different from the known template, and the matching residual is forced to be 0.8 to avoid misjudging the atypical noise as a pathological signal.

[0045] In the implementation of the application, the result fusion rule of the quality level output module is: the suspected pathological signals and noises determined by the pathological feature confidence calculation algorithm are retained, the weak pathological signals and pseudo-pathological noises output by the secondary classification algorithm are added, and four complete classifications are formed; among them, the weak pathological signal needs to be associated with the corresponding template type in the preset clinical pathological feature library to provide reference for clinical diagnosis.

[0046] When fusing, first count the proportion of each of the four types of signals, such as suspected pathological signal 28%, noise 15%, weak pathological signal 45%, and pseudo-pathological noise 12%. According to the fact that the proportion of weak pathological signal is the highest and exceeds 50%, the overall level is determined as “clinical observation is needed”; for each weak pathological signal, find the template with the highest matching degree (for example, a weak pathological signal has a matching degree of 92% with a ventricular premature beat template), and mark “weak pathological signal (suspected ventricular premature beat)” in the output result, so that the doctor can quickly understand the potential pathological type and improve the diagnosis efficiency.

[0047] While embodiments of the present application have been shown and described, it is to be understood that the embodiments described are merely divergences, modifications, replacements and variations of the embodiments, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for assessing the quality of electrocardiogram (ECG) signals based on unsupervised cascaded adaptive networks, characterized in that, Includes the following steps: S1. Preprocess the acquired raw ECG signal: remove power frequency interference by adaptive notch filtering, remove high-frequency electromyographic noise by soft threshold wavelet filtering, and retain potential pathological waveforms during the filtering process to obtain the preprocessed ECG signal. S2, Unsupervised Multimodal Feature Extraction: An unsupervised feature extraction network based on an autoencoder extracts time-domain features, frequency-domain features, and nonlinear features from the ECG signal preprocessed in S1, and then fuses the three types of features to generate a comprehensive feature vector. S3. Cascaded Adaptive Evaluation: Construct a two-level adaptive evaluation network to process the comprehensive feature vector, specifically including: First-level noise filtering: The improved K-means algorithm is used to cluster and filter significant noise features in the comprehensive feature vector to remove obvious noise signals and output candidate signals; Second-level pathology-noise differentiation: An adaptive attention mechanism is introduced, and the candidate signals output from the first level are initially classified by the pathological feature confidence calculation algorithm, and the initial quality assessment results are output. The initial quality assessment results include three categories: suspected pathological signals, noise, and ambiguous interval signals. At the same time, the ambiguous interval signals are further subdivided by a secondary classification algorithm, and the subdivided results are output. The subdivided results include two categories: weak pathological signals and pseudo-pathological noise. S4. Output Quality Level: Combining the preliminary quality assessment results and subdivision results, as well as the preset clinical pathology feature library, output complete ECG signal quality assessment results. The complete ECG signal quality assessment results include four categories: suspected pathological signals, weak pathological signals, pseudo-pathological noise, and noise.

2. The ECG signal quality assessment method based on unsupervised cascaded adaptive networks according to claim 1, characterized in that, The threshold calculation method for the soft threshold wavelet filter in the S1 ECG signal preprocessing module is as follows: based on the high-frequency coefficients after wavelet decomposition, the noise standard deviation is estimated, and combined with the number of sampling points of the original ECG signal, the threshold is determined through logarithmic relationship to ensure that high-frequency electromyographic noise with a frequency greater than 100Hz is removed without losing low-frequency pathological features.

3. The ECG signal quality assessment method based on unsupervised cascaded adaptive networks according to claim 1, characterized in that, The extraction method for time-domain features in the unsupervised multimodal feature extraction module of S2 is as follows: waveform segments of the preprocessed ECG signal are extracted using the sliding window method, and the QRS amplitude, RR interval standard deviation, and ST segment offset of each segment are calculated, totaling three time-domain parameters. The extraction method for frequency-domain features is as follows: fast Fourier transform is performed on the ECG signal preprocessed in S1, and the power spectral entropy and the energy ratio of the low-frequency and high-frequency bands of heart rate variability are calculated, totaling two frequency-domain parameters. The extraction method for nonlinear features is as follows: the fractal dimension is calculated using the box counting method, and the sample entropy is calculated using the sample entropy algorithm, totaling two nonlinear parameters. A total of seven parameters are fused with equal weights to generate a comprehensive feature vector.

4. The ECG signal quality assessment method based on unsupervised cascaded adaptive networks according to claim 1, characterized in that, The network structure parameters of the autoencoder in the unsupervised multimodal feature extraction module in S2 are as follows: the coding layer dimension is set to 128, the decoding layer dimension is consistent with the dimension of the preprocessed ECG signal, the mean square error is used as the loss function, and the distribution pattern of normal ECG signals is learned by minimizing the loss function, so as to provide a benchmark for the subsequent pathological feature confidence calculation algorithm.

5. The ECG signal quality assessment method based on unsupervised cascaded adaptive networks according to claim 1, characterized in that, The algorithm for calculating the confidence of pathological features in the second-level pathological-noise region molecular network of the cascaded adaptive evaluation module is used to achieve preliminary classification of candidate signals. The specific calculation formula includes: ； In the formula: : Confidence level of pathological features, with a value range of [0,1], used as a preliminary classification basis; when A value ≥0.6 is initially considered a suspected pathological signal. If the value is less than 0.3, it is initially determined to be noise; if 0.3 ≤ Signals with a value less than 0.6 are marked as fuzzy interval signals; The weight of the i-th pathological association feature includes three core pathological association features: ST segment offset, fractal dimension, and sample entropy, with corresponding weights of 0.4, 0.3, and 0.3, respectively. The weights are obtained by training with 500 myocardial infarction samples and 300 atrial fibrillation samples from a pre-set clinical pathological feature library. : No. The normalized value of each pathological correlation feature ranges from [0,1]. When the ST segment offset is greater than 0.1mV, the normalized value is 1. When the fractal dimension is less than 1.0, the normalized value is 1. When the sample entropy is less than 0.5, the normalized value is 1. : Reconstruction error correction coefficient, with a value of 0.2, is used to balance the influence of signal reconstruction results from unsupervised learning and the weights of pathological features; The reconstruction error of the autoencoder ranges from [0, 0.1]. The reconstruction error of the pathological waveform is higher than that of the normal signal but lower than that of the random noise due to deviation from the normal signal distribution. The reconstruction error of the normal signal is <0.02, the reconstruction error of the pathological signal is 0.02~0.05, and the reconstruction error of the noise is >0.

05.

6. The ECG signal quality assessment method based on unsupervised cascaded adaptive networks according to claim 5, characterized in that, Confidence of pathological features The initial classification threshold was determined by optimizing the receiver operating characteristic curve: using the pathological waveforms and noise marked in the preset clinical pathological feature library as the gold standard, the threshold was adjusted to make the overall accuracy of the initial classification ≥90%, while reserving a subdivision ratio of 10%~15% of the total signal for fuzzy interval signals.

7. The ECG signal quality assessment method based on unsupervised cascaded adaptive networks according to claim 1, characterized in that, The pre-defined clinical pathological feature database was constructed as follows: 1000 clinically diagnosed pathological electrocardiogram (ECG) samples and 500 normal ECG samples were collected. The pathological ECG samples included 500 myocardial infarction samples, 300 atrial fibrillation samples, and 200 premature ventricular contraction (PVC) samples. Waveform feature parameters were labeled for each sample, including QRS width, ST segment elevation amplitude, and RR interval regularity. Three typical pathological waveform templates were generated using K-means clustering. The parameters for the three typical pathological waveform templates were as follows: The first type of template corresponds to myocardial infarction, and its parameters are ST segment elevation amplitude ≥0.1mV in the precordial leads or ≥0.2mV in the limb leads, and the QRS wave is QS type; The second type of template corresponds to atrial fibrillation, with parameters including an absolutely irregular RR interval, an f-wave frequency of 350~600 beats / min, and an f-wave amplitude of <0.1mV; The third type of template corresponds to premature ventricular contractions (PVCs), with parameters including QRS width > 120 ms and complete compensatory pause. The pre-set clinical pathology feature library has been clinically validated by tertiary hospitals, ensuring that the matching accuracy between the template and the real pathological waveform is ≥95%.

8. The ECG signal quality assessment method based on unsupervised cascaded adaptive networks according to claim 5, characterized in that, The secondary classification algorithm of the second-level pathological-noise region molecular network in the cascaded adaptive evaluation module is used to subdivide the signal in the fuzzy region. The output subdivision results include two categories: weak pathological signals and pseudo-pathological noise. The specific calculation formula includes: ； In the formula: : Secondary classification confidence score, with a value range of [0,1], is used as the basis for subdivision; when A value ≥0.5 is considered a weak pathological signal. A value <0.5 is considered pseudo-pathological noise; The dispersion of pathological association features, with a value range of [0,1], is calculated as the ratio of the standard deviation of the three core pathological association features to the maximum value of each feature. : The matching residual between the signal and the preset clinical pathology feature library template, with a value range of [0,1]. It is calculated by the Euclidean distance between the comprehensive feature vector of the fuzzy interval signal and the feature vector of the most similar template in the preset clinical pathology feature library, and then normalized to the [0,1] interval. : The secondary classification balance coefficient, with a value of 0.4, is used to balance the weights of feature dispersion, matching residuals, and pathological feature confidence C; Pathological feature confidence score, which is the preliminary classification result output by the pathological feature confidence score calculation algorithm, with a value range of 0.3 ≤ <0.

6.

9. The ECG signal quality assessment method based on unsupervised cascaded adaptive networks according to claim 8, characterized in that, The selection rule for the most similar template in the secondary classification algorithm is as follows: For the comprehensive feature vector F_candidate of the fuzzy interval signal, calculate its Euclidean distance with the feature vectors of the three types of templates in the preset clinical pathology feature library, and select the template with the smallest distance as F_template; if the smallest Euclidean distance is greater than the preset threshold, which is 1.2 times the average Euclidean distance of all template feature vectors in the preset clinical pathology feature library, then e_match is forcibly set to 0.

8.

10. The method for assessing ECG signal quality based on an unsupervised cascaded adaptive network according to claim 9, characterized in that, The result fusion rule of the quality level output module is as follows: retain the suspected pathological signals and noise determined by the pathological feature confidence calculation algorithm, and add the weak pathological signals and pseudo-pathological noise output by the secondary classification algorithm to form four complete classifications; among them, the weak pathological signals need to be associated with the corresponding template type in the preset clinical pathological feature library to provide a reference for clinical diagnosis.

Images