CN-121971101-A - Multi-domain electrocardio intelligent analysis method of mixed Fourier and wavelet convolution neural network

CN121971101ACN 121971101 ACN121971101 ACN 121971101ACN-121971101-A

Abstract

The invention provides a multi-domain electrocardio intelligent analysis method of a mixed Fourier and wavelet convolution neural network, belongs to the technical field of intersection of artificial intelligence and medical health, and solves the technical problems that the traditional single-domain ECG analysis is difficult to consider global spectrum characteristics and local transient characteristics and poor in task generalization. The technical scheme includes that the method comprises the following steps of S1, preprocessing an ECG signal, filtering, segmentation, normalization and data enhancement, S2, constructing a three-branch model, S3, fusing multi-domain features by an attention mechanism, S4, setting a task classification head, S5, optimizing by Adam and preventing overfitting by an early-stop strategy, S6, inputting data, outputting arrhythmia classification, biological identification and sleep apnea detection results. The multi-domain fusion of the invention improves generalization and accuracy, supports multitasking, and adapts to clinical and biological recognition scenes.

Inventors

GAO ZHAN
WANG SAIYU
Asif Nurul Hakim
JU XIAOLIN
PAN HAIYAN
SHAO HAORAN

Assignees

南通大学

Dates

Publication Date: 20260505
Application Date: 20251118

Claims (9)

1. The multi-domain electrocardio intelligent analysis method of the mixed Fourier and wavelet convolution neural network is characterized by comprising the following steps of: S1, acquiring and preprocessing a multisource electrocardiograph data set, wherein the data set comprises an arrhythmia data set MIT-BIH, a biological identification data set ECG-ID and a sleep Apnea data set Apnea-ECG; S2, dividing data sets, namely an arrhythmia data set MIT-BIH and a sleep Apnea data set Apnea-ECG, respectively dividing the two data sets into training sets, verification sets and test sets according to the proportion of 70 percent to 15 percent and 15 percent; S3, constructing an HFW-CNN three-branch model comprising a time domain branch, a Fourier domain branch and a wavelet domain branch; S4, attention guiding feature fusion; s5, configuring a multi-task classification head; S6, training a model: adam optimizer, the learning rate is reduced by 10 times when the verification loss is stopped; early stopping, namely stopping the continuous 15 rounds of verification performance without lifting, and reserving optimal parameters; regularization measures, namely, inhibiting overfitting by combining He normal initialization, L2 weight attenuation and Dropout; Implementing configuration, namely implementing based on PyTorch 2.0.0 frames and adopting NVIDIA RTX 4090GPU to accelerate training and reasoning; And S7, outputting a multitasking result, namely outputting arrhythmia category and confidence level, biological recognition subject ID and matching degree, sleep apnea detection tag and probability, and supporting front-end visual display.
2. The multi-domain electrocardiographic intelligent analysis method of hybrid fourier and wavelet convolutional neural network according to claim 1, wherein the S1 preprocessing further comprises the steps of: S11, analyzing data and unifying formats, carrying out layered filtering, denoising and correcting, combing and marking associated signals, removing unqualified fragments by quality control, dividing the samples according to 70 percent to 15 percent after sample equalization, and providing reliable data for subsequent training when the samples do not cross sets at night; S12, dividing the sleep apnea detection device into 60S of non-overlapping epochs, namely taking a 100Hz sampling rate as a reference, wherein each 60S corresponds to 6000 sampling points, continuously intercepting the sleep apnea detection device according to a time axis from a beginning end of night recording, and accurately aligning each epoch with a minute-level respiratory event mark, directly removing a residual section with the tail less than 60S, ensuring the integrity of each epoch signal, and adapting to the feature extraction requirement of subsequent sleep apnea detection; S13, labeling "Apnea" and "normal", wherein expert labeling files matched with an Apnea-ECG data set are firstly extracted, a minute-level respiratory event label is contained, effective labeling of "Apnea" and "normal" is screened out, labeling missing "and" invalid "items are removed, labeling of each minute is accurately aligned with corresponding 60S non-overlapping epochs according to a time axis, 1 epoch is bound with 1 label, finally, the labeling is converted into a model adaptation format, and A=" Apnea "and N=" normal ", so that the labeling corresponds to signal fragments one by one, and clinical sleep Apnea diagnosis standard is met.
3. The multi-domain electrocardiographic intelligent analysis method of hybrid fourier and wavelet convolutional neural network according to claim 1, wherein S3 further comprises the steps of: S31, fourier domain branching, calculating amplitude spectrum As the real part of the component, Is an imaginary part; s32, inputting an amplitude spectrum into the 1D-CNN, and advancing global spectrum characteristics; And S33, wavelet domain branching, namely performing continuous wavelet transformation CWT on each epoch to generate a 2D time-frequency wavelet scale map scalogram S (a, b) = |W x (a,b)| 2 ,W x (a, b) as a CWT result, W x (a, b) as coefficients of continuous wavelet transformation, a as a scale factor and b as a translation factor, and inputting scalogram into the 2D-CNN to extract local transient abnormal characteristics.
4. The multi-domain electrocardiographic intelligent analysis method of hybrid fourier and wavelet convolutional neural network according to claim 1, wherein S4 further comprises the steps of: s41, performing linear transformation on the spliced characteristic F concat ＝[F time ||F Fourier ||F wavelet to obtain a time domain characteristic of Z=W a F concat +b a ,F time , a Fourier domain characteristic of F Fourier , a wavelet domain characteristic of F wavelet , a weight matrix capable of being learned of W a , and a characteristic dimension of d concat ×d concat ,d concat and F concat ; S42. the dimension of the attention weight α=softmax (Z) is consistent with F concat ; S43, fusion characteristic batch normalization, wherein BatchNorm operation is carried out on the fusion characteristic F fusion ＝α⊙F concat , and the steps are as follows For normalized feature tensor, e=1e-5, μ F is the batch mean, σ is the batch standard deviation, and then pass The scaling and translation are completed, F bn is the final output characteristic after BatchNorm, and gamma and beta are the learnable parameters.
5. The multi-domain electrocardiographic intelligent analysis method of hybrid fourier and wavelet convolutional neural network according to claim 1, wherein S6 further comprises the steps of: The method comprises the steps of S61, assigning weights according to the reciprocal of the class sample duty ratio, wherein the lower the class duty ratio is, the higher the weights are, and the biological recognition ECG-ID adopts a common cross entropy Loss function; And S62, monitoring strategies, namely calculating core performance indexes of the verification set after each round of training, namely weighing F1-score of arrhythmia and ROC-AUC (ROC-AUC) detected by apnea, and terminating training and storing model parameters with optimal comprehensive indexes when the indexes are not improved for 15 consecutive rounds.
6. The multi-domain electrocardiographic intelligent analysis method of hybrid fourier and wavelet convolutional neural network according to claim 1, wherein S7 further comprises the steps of: S71, performance, namely arrhythmia accuracy rate is more than or equal to 98.0%, F1 is more than or equal to 0.65, biological identification accuracy rate is more than or equal to 97.1%, F1 is more than or equal to 0.97%, apnea detection accuracy rate is more than or equal to 95.0%, and AUC is more than or equal to 0.98; And S72, visualizing a front-end display waveform diagram, and marking key cardiac epoch fragments, diagnosis labels, biological recognition matching degree and auxiliary result interpretation by the associated classification results.
7. The multi-domain electrocardiographic intelligent analysis method of mixed fourier and wavelet convolutional neural network according to claim 1, wherein in S3, constructing the HFW-CNN three-branch model comprises: Time domain branching, namely adopting 1D-CNN, convolution kernel sizes being 7, 5 and 3, the number of filters being 64, 128 and 256 in sequence, including BatchNorm and Dropout, and discarding rate being 0.3-0.5, for extracting morphological characteristics of P wave, QRS wave and T wave; fourier domain branching, namely performing fast Fourier transform on the signal to obtain a frequency domain spectrum, and then accessing 1D-CNN, wherein the convolution kernel size is 5 and 3, and the number of filters is 64 and 128 so as to capture global spectrum characteristics; And (3) branching a wavelet domain, namely executing continuous wavelet transformation on each epoch, inputting scalogram obtained by the continuous wavelet transformation into 2D-CNN, and realizing positioning of ventricular premature beat and apnea related pause local transient abnormality with the number of filters being 64 and 128.
8. The multi-domain electrocardio intelligent analysis method of the mixed Fourier and wavelet convolution neural network is characterized in that attention guiding characteristics of S4 are fused specifically by firstly splicing time domain characteristics, fourier domain characteristics and wavelet domain characteristics, obtaining weights through calculation containing learnable parameters, performing element multiplication with the spliced characteristics after being activated by softmax to complete fusion, performing BatchNorm operation after fusion, normalizing the mean and variance of the fused characteristics, and introducing learnable scaling and translation parameters to obtain final fused characteristics.
9. The method for intelligent analysis of multi-domain electrocardio of a hybrid Fourier and wavelet convolutional neural network of claim 1, wherein the S5 configuration multi-task classification head comprises an arrhythmia classification MIT-BIH, wherein a Softmax classification head is adopted to output 5 types of labels, namely normal Dou Lv N, ventricular premature beat V, supraventricular premature beat S, fusion wave F or O, rare arrhythmia and corresponding prediction confidence; The biological recognition ECG-ID adopts a Softmax classification head to output 90 classes of subject IDs and matching confidence; Sleep Apnea detection Apnea-ECG, wherein a Sigmoid classification head is adopted to output 'Apnea' and 'normal' classification labels and prediction probability.

Description

Multi-domain electrocardio intelligent analysis method of mixed Fourier and wavelet convolution neural network Technical Field The invention relates to the technical field of intersection of artificial intelligence and medical health, in particular to a multi-domain electrocardio intelligent analysis method of a hybrid Fourier and wavelet convolutional neural network. Background Cardiovascular diseases are the leading cause of death worldwide, sleep apnea is used as a common sleep disorder and can induce cardiovascular complications, early screening and intervention of the cardiovascular diseases and the sleep apnea are important to reduce disability rate and death rate, and simultaneously, electrocardiography is a core carrier for high-safety biological recognition due to the fact that the electrocardiography contains individual specific physiological characteristics. ECG is a non-invasive, low-cost and highly real-time technique, and is widely used in clinical arrhythmia diagnosis, sleep apnea detection and biological recognition, and particularly, along with popularization of wearable equipment and telemedicine, the demand for high-precision and multi-scene intelligent analysis of ECG signals has been rapidly increased. The traditional ECG analysis method is highly dependent on the experience of medical specialists or is realized by combining manual design characteristics such as RR interval and QRS wave width with a rule algorithm, but has obvious limitations that on one hand, ECG signals have the characteristics of weak waveforms, easiness in being influenced by myoelectric interference, power frequency noise and baseline drift, and the inter-individual waveforms have large differences and complex forms, so that the robustness of manual characteristic extraction is poor, the generalization capability is weak, and the large-scale automatic analysis is difficult to adapt, and on the other hand, the method can only capture single dimension information, can not compromise the morphological characteristics, global spectrum rule and local transient abnormality of electrocardiosignals, and has insufficient accuracy in identifying complex diseases such as rare arrhythmia and sleep apnea related signal disturbance. In recent years, the deep learning technology provides a new direction for ECG analysis, namely a convolutional neural network CNN is good at extracting local morphological characteristics, a long-short-term memory network LSTM can model time sequence dependence, a attention mechanism can focus a key wave band, but the existing method still has three major core bottlenecks, namely, one single domain represents defects that a majority of models only depend on a single signal domain, fourier transformation can capture global periodicity but lacks time positioning capability, wavelet transformation can identify local time-frequency anomalies but lose global spectrum context, time domain analysis only focuses on waveform morphology and ignores frequency domain information, so that multidimensional characteristics of an ECG signal cannot be comprehensively described, the second task specificity is limited, the existing models are designed for single tasks, such as arrhythmia classification or single biological recognition cannot be simultaneously adapted, electrocardio biological recognition and sleep apnea detection, clinical deployment needs multi-model parallelism, hardware cost is high and compatibility is poor, robustness and generalization are low, tolerance to noise and individual difference is low, the multi-dimensional characteristics cannot be fully verified based on a single data set, the multi-functional classification and biological recognition performance is not covered, the practical requirements and the practical requirements of the multi-functional classification and the biological recognition scene are high, and the clinical requirements of the clinical requirements are difficult to identify. How to break through the limitation of single-domain analysis, realize multi-task adaptation, and improve the robustness and generalization capability of ECG intelligent analysis becomes a technical problem to be solved in the invention. Disclosure of Invention The invention aims to provide a multi-domain electrocardio intelligent analysis method of a mixed Fourier and wavelet convolution neural network, which can break through the limitation of single-domain analysis and the bottleneck of single-task adaptation, realize arrhythmia classification, electrocardio biological identification and sleep apnea detection, and improve the accuracy, robustness and generalization capability of ECG analysis under multiple scenes. The invention provides an HFW-CNN architecture of three-branch parallel feature extraction and attention guidance fusion, which aims at the requirements of 'non-stationary, multidimensional, noise interference prone' and multitasking of ECG signals, and combines the characteristic desig