ECG high frequency feature analysis method for myocardial ischemia and complex heart rate abnormalities

Abstract

The application discloses an ECG high-frequency feature analysis method for myocardial ischemia and complex heart rate abnormalities, and belongs to the technical field of medical information processing and artificial intelligence, comprising: obtaining an ECG signal and performing preliminary classification based on a support vector machine; adopting a random forest model of a fusion model irrelevant meta-learning and a domain self-adaptive mechanism to dynamically compensate individual physiological differences; generating an individualized dynamic diagnosis threshold sequence through reinforcement learning; mining potential abnormal misjudgment modes by using a variational autoencoder and a generative adversarial network; fusing features and clarifying abnormal state boundaries by means of a graph neural network; and finally realizing gradient quantitative evaluation of the severity of myocardial ischemia through a multi-task learning model; the application effectively solves the problems of poor individual adaptability, fixed threshold and insufficient recognition of complex abnormal patterns of traditional methods, and significantly improves the accuracy, stability and clinical practicability of electrocardiogram analysis in the diagnosis of myocardial ischemia and complex heart rate abnormalities.

Description

Technical Field

[0001] This invention belongs to the field of medical information processing and artificial intelligence technology, and in particular relates to a method for ECG high-frequency feature analysis for myocardial ischemia and complex heart rate abnormalities. Background Technology

[0002] In the field of medical diagnostics, high-frequency feature analysis of electrocardiogram (ECG) signals plays a crucial role in identifying myocardial ischemia and complex arrhythmias. This research direction is directly related to the early detection and timely intervention of cardiovascular diseases, and has an undeniable significance for improving patient survival rates and quality of life. With technological advancements, more and more research focuses on capturing subtle changes in ECG signals to aid in diagnosis, but this field still faces many challenges and urgently needs breakthroughs.

[0003] Currently, although many methods can detect heart rate abnormalities using electrocardiogram (ECG) data, most approaches often struggle to accurately adapt to the specific circumstances of different populations, especially when dealing with individual differences and complex medical conditions. In particular, when analyzing heart rate changes, existing technologies focus more on signal characteristics under fixed standards, neglecting the influence of the patient's own background on heart rate performance. This neglect leads to inconsistent diagnostic accuracy, especially when dealing with diverse patient populations, resulting in frequent misdiagnosis or missed diagnosis.

[0004] More importantly, this field still faces significant technical challenges, particularly in scientifically defining the criteria for judging heart rate changes. Heart rate fluctuations naturally vary from person to person; without dynamic adjustments based on each patient's specific situation, it's difficult to accurately distinguish between normal and abnormal states. A deeper problem is that this lack of clear criteria further impacts the identification of different degrees of myocardial ischemia. For example, in some cases, minor abnormalities may be overlooked, while severe abnormalities may be misjudged as normal fluctuations. This escalating challenge makes the diagnostic process fraught with uncertainty. For instance, when monitoring an elderly patient with a low baseline heart rate, simply using general standards might misinterpret normal fluctuations as abnormalities, leading to unnecessary medical interventions.

[0005] Therefore, how to construct a flexible and accurate standard for judging heart rate changes in different individuals in high-frequency electrocardiogram feature analysis, and how to accurately identify the characteristic manifestations of myocardial ischemia through this standard, has become a key problem that current research urgently needs to overcome. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention provides a method for ECG high-frequency feature analysis of myocardial ischemia and complex heart rate abnormalities; by constructing a multi-stage, algorithm-integrated intelligent processing pipeline, it significantly improves the accuracy of diagnosis, individual adaptability and clinical applicability.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The ECG high-frequency feature analysis method for myocardial ischemia and complex heart rate abnormalities specifically includes the following steps: S1. Acquire electrocardiogram (ECG) signal data, extract relevant raw records containing high-frequency feature analysis from a pre-established patient database, perform preliminary classification of the raw records using a support vector machine, and obtain pre-classified ECG signal data groups. S2. For the electrocardiogram signal data grouping after the preliminary classification, compensation processing is performed to ignore individual differences. The compensation-processed electrocardiogram signal data grouping is determined by calculating the weight distribution of background variables of each patient in the group. S3. Based on the ECG signal data grouping after compensation processing, obtain the relevant fluctuation index of unstable diagnostic accuracy. If the fluctuation index exceeds the preset threshold, apply the filtering algorithm to the grouping to adjust the boundary and obtain the ECG signal data grouping after boundary adjustment. S4. Based on the ECG signal data grouping after the boundary adjustment, determine the dynamic requirements set by the judgment criteria. If the coefficient of variation of the heart rate change adjustment shown in the grouping is higher than the threshold, generate a dynamic threshold sequence and determine the ECG signal data grouping after the application of the dynamic threshold sequence. S5. Extract potential patterns of misjudgment of normal fluctuations from the ECG signal data group after the application of the dynamic threshold sequence, and judge the abnormal tendency of the potential patterns by comparing the similarity between the group and the historical benchmark data, and obtain the ECG signal data group after the abnormal tendency judgment. S6. For the ECG signal data grouping after the abnormal tendency judgment, analyze the intersection of the abnormal state boundary. If the intersection is located within the preset boundary interval, fuse the feature vector within the boundary interval to determine the ECG signal data grouping after fusing the feature vector. S7. Based on the grouping of ECG signal data after fusion feature vectors, evaluate the gradient distribution of severity identification, and obtain the ECG signal data grouping after quantization scale mapping by mapping the grouping to the quantization scale identification of myocardial ischemia.

[0008] As a further preferred embodiment of the ECG high-frequency feature analysis method for myocardial ischemia and complex heart rate abnormalities of the present invention, step S1 specifically includes the following sub-steps: acquiring ECG signal data and extracting original records from a pre-established patient database; extracting high-frequency features based on the original records to obtain high-frequency feature data; classifying the high-frequency feature data using a support vector machine to obtain classified ECG signal data groups; calculating heart rate variability indices for the classified ECG signal data groups to determine the initial distribution of heart rate abnormalities; if the heart rate variability index exceeds a preset threshold, marking the corresponding group as an abnormal group to obtain abnormally marked ECG signal data; performing cluster analysis based on the abnormally marked ECG signal data to determine sub-groups within the abnormal group; extracting heart rate abnormality patterns from the sub-groups to obtain heart rate abnormality pattern data.

[0009] As a further preferred embodiment of the ECG high-frequency feature analysis method for myocardial ischemia and complex heart rate abnormalities of the present invention, step S2 specifically includes the following sub-steps: ranking the background variable values ​​by feature importance using random forest to determine the main background variable set; calculating the similarity matrix among patients within the group for the main background variable set to obtain similarity distribution data; adjusting the weight allocation of the compensated ECG signal data groupings according to the similarity distribution data to determine the weighted ECG signal data groupings; extracting time-domain heart rate variability parameters from the weighted ECG signal data groupings to obtain time-domain parameter data; if the time-domain parameter data exceeds a preset threshold range, marking the corresponding group as a potential abnormal group to determine the ECG signal data groupings with potential abnormality markings; performing frequency-domain heart rate variability analysis on the ECG signal data groupings with potential abnormality markings to obtain frequency-domain parameter data.

[0010] As a further preferred embodiment of the ECG high-frequency feature analysis method for myocardial ischemia and complex heart rate abnormalities of the present invention, step S3 specifically includes the following sub-steps: Processing the ECG signal data into groups to obtain fluctuation index data of diagnostic accuracy within each group, and determining the specific distribution of the fluctuation index; based on the distribution of the fluctuation index, determining whether it exceeds a preset threshold range; if it exceeds the preset threshold range, applying a filtering algorithm to the group boundaries of the data groups to obtain boundary-adjusted data groups; for the boundary-adjusted data groups, extracting signal processing features within each group, analyzing the correlation between features and data stability, and determining the stability parameters of the feature distribution; based on the stability parameters of the feature distribution, obtaining the signal processing parameters within each group... If the consistency level of the processed data is lower than the preset standard, a second smoothing process is performed on the signals within the group to obtain data groups with improved consistency. For the data groups with improved consistency, the changing trend of ECG signals between different groups is analyzed to determine whether the changing trend meets the preset stationarity conditions. If not, the data between groups is equalized to obtain equalized data groups. Through the equalized data groups, the index analysis results within each group are extracted, and combined with the requirements of diagnostic accuracy, the stability assessment data of the final group is determined. Based on the stability assessment data of the final group, a signal processing report for each group is generated to determine whether the signal processing has achieved the preset stability target, thus obtaining the final processed ECG signal data groups.

[0011] As a further preferred embodiment of the ECG high-frequency feature analysis method for myocardial ischemia and complex heart rate abnormalities of the present invention, S4 specifically includes the following sub-steps: by grouping the data of the ECG signal, obtaining the variation data of the heart rate change of each group, and performing preliminary screening on the variation data to obtain a set of groups that exceed a preset threshold; For the selected group set, a dynamic threshold generation method is used to construct a threshold sequence matching the individual range, determining the generated dynamic threshold sequence data. Based on the generated dynamic threshold sequence data, it is applied to the corresponding data group to update the ECG signal grouping structure, resulting in a processed signal group set. For the processed signal group set, the matching degree of signals within each group is analyzed. If the matching degree is lower than a preset standard, the signals within the group are smoothed to obtain more consistent group data. Using the more consistent group data, the transition of heart rate changes between groups is obtained. If the transition shows discontinuity, interpolation is performed on the signals between groups to determine the smoothed transition group data. Based on the smoothed transition group data, signal stability parameters within each group are extracted, and it is determined whether the stability parameters meet preset requirements, resulting in the final processed ECG signal group data. For the final processed ECG signal group data, the correspondence between the individual range and the dynamic threshold is analyzed to determine the threshold application results for each group.

[0012] As a further preferred embodiment of the ECG high-frequency feature analysis method for myocardial ischemia and complex heart rate abnormalities of the present invention, step S5 specifically includes the following sub-steps: obtaining relevant information on fluctuation extraction from the ECG signal data grouping after applying a dynamic threshold sequence; analyzing the potential hidden patterns in the fluctuation extraction results to determine the basic structure of the potential patterns; based on the basic structure of the potential patterns, using a pre-established classification model, specifically using the support vector machine method, to perform preliminary classification of the potential patterns, obtaining a set of classified patterns; comparing the classified pattern set with historical benchmark data to analyze the matching situation between the pattern set and the historical benchmark data; if the matching degree is lower than a preset threshold, it is marked as an abnormal tendency, obtaining a set of marked patterns; from the marked... From the pattern set, a subset of patterns with abnormal tendencies is extracted. For this subset, its distribution in signal groups is analyzed to determine the specific location of the abnormal tendencies within the groups. Based on the specific location of the abnormal tendencies within the groups, the classification boundaries of the signal groups are adjusted. For the adjusted groups, the fluctuation characteristics within each group are recalculated to obtain updated signal groups. Through the updated signal groups, the distribution of misjudgment risk within each group is analyzed. If the misjudgment risk in a certain group is higher than a preset standard, the signal data in that group is weighted to determine the final signal groups that distinguish misjudgment risks. For the final signal groups that distinguish misjudgment risks, the relevant information on the abnormal tendencies and fluctuation extraction within each group is recorded to generate a structured group file, resulting in signal data groups that can be used for subsequent analysis.

[0013] As a further preferred embodiment of the ECG high-frequency feature analysis method for myocardial ischemia and complex heart rate abnormalities of the present invention, in step S6, the feature vectors within the fusion boundary interval to enhance boundary clarity include: ECG feature vectors near the boundary interval are constructed into a feature graph, where nodes represent feature vectors and edges represent the similarity or temporal correlation strength between features. A graph attention network is applied to aggregate information from neighboring nodes and update the feature representation of each node, thereby enabling features of the same type of abnormality to cluster in space and features of different categories to separate from each other. Spectral clustering analysis is performed on the updated node features to redefine the classification boundaries of abnormal states based on the distribution structure of the data itself. The final grouping with improved boundary clarity is output, where each group corresponds to a heart rhythm or myocardial ischemia state with a clearly defined feature.

[0014] As a further preferred embodiment of the ECG high-frequency feature analysis method for myocardial ischemia and complex heart rate abnormalities of the present invention, step S7 specifically includes the following sub-steps: For the ECG signal data grouped after fusion of feature vectors, obtain the signal data fluctuation characteristics within each group, analyze the gradient distribution of the fluctuation characteristics in severity, and obtain the gradient-distributed signal data groupings; based on the gradient-distributed signal data groupings, extract signal feature points related to myocardial ischemia within each group, and use a pre-established support vector machine model to classify the feature points, determining the classified signal feature point set; for the classified signal feature point set, analyze the distribution position of the feature points on the quantization scale; if the distribution position deviates from the preset threshold range, reclassify the deviated feature points to obtain the reclassified and adjusted signal feature point set; through... The system categorizes and adjusts the set of signal feature points, obtains the correlation strength between each feature point and myocardial ischemia identification, analyzes the distribution pattern of the correlation strength on the quantization scale, and determines the set of signal feature points after the correlation strength distribution. Based on the set of signal feature points after the correlation strength distribution, the system performs scale transformation on the signal data within each group to generate scale-transformed ECG signal data groups. For the scale-transformed ECG signal data groups, the system analyzes the matching degree between the signal features within each group and myocardial ischemia. If the matching degree is lower than a preset threshold, the signal features within the group are redistributed to obtain the final adjusted ECG signal data groups. Through the final adjusted ECG signal data groups, the system obtains the distribution consistency of the signal features within each group, judges whether the consistency meets the preset standard, and generates the final ECG signal data groups that meet the standard. A system based on an ECG high-frequency feature analysis method for myocardial ischemia and complex heart rate abnormalities includes: a data acquisition module for acquiring ECG signal data and extracting raw records from a patient database; a preliminary classification module for performing preliminary classification of the raw records using a support vector machine to obtain preliminarily classified ECG signal data groups; an individual difference compensation module for using a random forest to compensate for neglected individual differences and determine the compensated groups; a boundary adjustment module for applying a filtering algorithm to adjust the boundaries of the groups based on fluctuation indices to obtain the adjusted boundary groups; a dynamic threshold generation module for generating a dynamic threshold sequence and applying it to the groups; and a misjudgment pattern recognition module for extracting potential patterns of misjudgment of normal fluctuations and judging abnormal tendencies. The feature fusion module is used to fuse feature vectors within the boundary range to enhance the clarity of the boundary; the severity quantization module is used to evaluate the gradient distribution of severity identification and map it to a quantization scale.

[0015] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects: This invention discloses a precise diagnostic method for myocardial ischemia based on electrocardiogram (ECG) signal processing. It extracts raw ECG records containing high-frequency features from a patient database, performs initial classification using a support vector machine (SVM) to form groups reflecting the initial distribution of abnormal heart rate. Subsequently, a random forest is used to compensate for individual differences, and the group boundaries are adjusted through weight distribution to improve stability. If the fluctuation index exceeds a threshold, a filtering algorithm is applied to optimize the boundaries, and a dynamic threshold sequence is generated based on the coefficient of variation to adapt to specific individual ranges. Potential misjudgment patterns are further extracted and compared with historical benchmarks to distinguish abnormal tendencies. Finally, the cross-point fusion feature vector is analyzed and mapped to a quantization scale to achieve precise differentiation of myocardial ischemia features. This invention significantly improves the accuracy and stability of myocardial ischemia diagnosis by progressively compensating for individual factors, dynamically adjusting thresholds, reducing the risk of misjudgment, and enhancing boundary clarity. Attached Figure Description

[0016] Figure 1 This is a flowchart of the ECG high-frequency feature analysis method for myocardial ischemia and complex heart rate abnormalities according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0018] like Figure 1 As shown, the ECG high-frequency feature analysis method of the present invention for myocardial ischemia and complex heart rate abnormalities may specifically include: S101. Obtain the electrocardiogram signal data, extract relevant original records containing high-frequency feature analysis from a pre-established patient database, perform preliminary classification on the original records using a support vector machine, and obtain preliminary classified electrocardiogram signal data groups, wherein the groups reflect the initial distribution of heart rate abnormality detection.

[0019] Acquire electrocardiogram (ECG) signal data by extracting raw records from a pre-established patient database. High-frequency features are extracted from the raw records to obtain high-frequency feature data. This high-frequency feature data is then classified using a support vector machine (SVM) to obtain grouped ECG signal data. Heart rate variability (HRV) indices are calculated for each group to determine the initial distribution of heart rate abnormalities. If the HRV index exceeds a preset threshold, the corresponding group is marked as an abnormal group, resulting in abnormally labeled ECG signal data. Cluster analysis is performed on the abnormally labeled ECG signal data to determine subgroups within each abnormal group. Heart rate abnormality patterns are extracted from each subgroup to obtain heart rate abnormality pattern data.

[0020] The specific calculations for acquiring electrocardiogram signal data and performing preliminary classification are as follows: (8) A one-dimensional convolutional neural network is used to automatically extract features from the original ECG signal, wherein the output features of the convolutional layer are... The specific calculations are as follows: ; in, Indicates the first The input ECG signal of each channel, where C is the total number of leads; For the first Layer The convolution kernel pairs with the first... One channel, This represents a one-dimensional convolution operation. For bias terms, It is the ReLU activation function; (2) A multi-head self-attention mechanism is introduced to enhance the attention to key temporal features. The specific calculation of the attention mechanism is as follows: in, , , , The feature matrix extracted by CNN, For learnable weight matrix, The feature dimension scaling factor; (3) The deep features are fused with traditional heart rate variability statistical features and then input into the support vector machine classifier. The specific calculation is as follows: ; Among them, among them, For support vector weights, For category labels, For kernel function, This is a bias term.

[0021] In the overall process of detecting abnormal electrocardiogram (ECG) signals, raw ECG signal data is first obtained from a pre-established patient database. For example...

[0022] In one possible implementation, the database stores 24-hour Holter monitoring records from multiple patients, each record containing sampling data from multiple leads, typically at a sampling rate of 250 Hz or higher. These raw records serve directly as the basis for subsequent analysis, ensuring the data source is reliable and complete.

[0023] Specifically, high-frequency feature extraction is one of the key steps. High-frequency features mainly refer to the high-frequency components in the QRS complex, such as the slope changes of the QRS wave or morphological details such as notched and slurred.

[0024] In one embodiment, the original signal is decomposed using wavelet transform to extract detail coefficients in the 4-40Hz range. These coefficients are sensitive to subtle abnormalities during ventricular depolarization. This extraction method helps to highlight subtle pathological changes that are difficult to capture with traditional low-pass filtering, thus providing more discriminative features for subsequent classification.

[0025] For example, when classifying high-frequency feature data using support vector machines, the feature vectors can be input into a pre-trained SVM model, which has been trained beforehand using labeled normal and abnormal samples.

[0026] In one possible implementation, the classification results divide the electrocardiogram signals into a normal group and a potentially abnormal group.

[0027] Preferably, SVM using radial basis function kernels can better handle nonlinearly separable feature distributions, improving classification accuracy. The beneficial effects of this classification are rapid screening of large amounts of data, reduced manual interpretation burden, and laying the foundation for subsequent targeted analysis. For potentially abnormal groups after classification, calculating heart rate variability indices is a crucial step in further confirming abnormalities. Heart rate variability indices include time-domain indices such as SDNN and RMSSD, and frequency-domain indices such as the LF / HF ratio. For example...

[0028] In one embodiment, the RR interval sequence is calculated for signals within the abnormal group. If the SDNN is less than 50 ms or the RMSSD is less than 15 ms, the heart rate variability is considered significantly reduced, and the group is labeled as the abnormal group. This threshold setting is based on clinical guidelines and can effectively identify heart rhythm abnormalities caused by autonomic dysfunction. After labeling the abnormal group, the benefit is that the focus is concentrated on high-risk data, avoiding over-analysis of normal signals.

[0029] It should be noted that cluster analysis based on abnormally labeled electrocardiogram signal data can further divide the abnormal groups into subgroups.

[0030] In one possible implementation, K-means clustering is used to unsupervisedly group the feature vectors of abnormal signals, for example, into subgroups of autonomic nervous system-related abnormalities and ischemia-related abnormalities. After clustering, each subgroup exhibits a more consistent abnormal pattern. Finally, heart rate abnormality patterns are extracted from the subgroups, such as the ST-segment elevation accompanied by accelerated heart rate commonly seen in the ischemia-related subgroup, or the alternation of long and short RR intervals appearing in the autonomic nervous system disorder subgroup. The extraction of these pattern data provides intuitive evidence for clinical diagnosis, with beneficial effects including improving the accuracy of fine-grained identification of abnormality types, supporting personalized treatment planning, and significantly improving the early warning capability of cardiovascular events.

[0031] S102. For the electrocardiogram signal data grouping after the preliminary classification, random forest is used to compensate for the neglect of individual differences. By calculating the weight distribution of background variables of each patient in the group, the compensated electrocardiogram signal data grouping is determined, wherein the compensated grouping integrates individual difference factors.

[0032] Based on the compensated ECG signal data grouping, the specific values ​​of background variables for each patient within each group are obtained. Random forest is used to rank the background variable values ​​by feature importance, determining the set of principal background variables. A similarity matrix is ​​calculated among patients within each group for the set of principal background variables, yielding similarity distribution data. The weight allocation of the compensated ECG signal data groupings is adjusted based on the similarity distribution data, determining the weighted ECG signal data groupings. Time-domain heart rate variability parameters are extracted from the weighted ECG signal data groupings, obtaining time-domain parameter data. If the time-domain parameter data exceeds a preset threshold range, the corresponding group is marked as a potentially abnormal group, determining the ECG signal data groupings with potentially abnormal markings. Frequency-domain heart rate variability analysis is performed on the ECG signal data groupings with potentially abnormal markings, obtaining frequency-domain parameter data.

[0033] Specifically, compensating for individual differences includes: (1) Apply the model-independent meta-learning framework for individual difference adaptive compensation, where the inner layer update is adapted to specific patients to allow the model to quickly adapt to the uniqueness of individual patients. The specific calculation is as follows: in, The base model consists of the learnable parameters of a neural network that has been pre-trained on a large amount of general data. For the first indivual The patient's learning task consists of a small amount of the patient's electrocardiogram data and its corresponding labels; The inner learning rate controls the step size when the model adjusts parameters to adapt to an individual patient; For loss function Regarding the basic model The gradient, the loss in the patient On the task of calculation, To adapt the model parameters, i.e., in the base model Based on the patient The new parameters are obtained after several steps of fast gradient update on the data; (2) The outer layer update achieves the meta-learning objective and is used to optimize the base model. : in, To optimize the objective, To distribute tasks from all patients The task sampled from different patients is to sum the results. Use the adapted model In patients The loss calculated on the task; (3) The total loss function of the domain adaptive random forest, used to reduce the differences in data distribution between different patient groups, is as follows: in, Total training loss; Cross-entropy loss is used for classification and is used to ensure that the model can accurately identify anomalies. To weigh the parameters, which are used to balance the importance of classification loss and domain loss; Domain loss is used to measure the difference between the source domain (training patient population) and the target domain (new patient feature distribution). Domain loss is measured using the maximum mean difference: in, For the first Each source domain is an existing, labeled training database sample; No. The target domain is the newly arrived patient sample that needs to be analyzed; These represent the number of samples in the source domain and the target domain, respectively. The feature mapping function maps the original data to the reproducing kernel Hilbert space; For the regenerating nucleus Hilbert space; The norm in the RKHS space is the distance metric; S25, Generate patient-specific weights based on meta-learning results : in, This is a temperature parameter used to control the sharpness of the weight distribution; This is a similarity measurement function; for responded to the patient The subsequent personalized model; In response to the patient The subsequent personalized model; Basic meta-model; This refers to the total number of patients.

[0034] For example, when processing compensated electrocardiogram (ECG) signal data grouping, it is first necessary to obtain the specific values ​​of background variables for each patient within the group. Background variables may include information such as age, gender, and past medical history, which are of significant reference value for the interpretation of ECG signals.

[0035] Specifically, assuming a group contains 10 patients aged 40 to 75, some of whom have a history of hypertension, this data will be compiled into a structured table for subsequent analysis.

[0036] For example, ranking the feature importance of background variable values ​​can be achieved using a random forest model. Random forests output importance scores by evaluating the influence of each variable on the classification result. Assuming that age and a history of hypertension have importance scores of 0.35 and 0.28, respectively, significantly higher than other variables, these two are identified as the set of primary background variables. This approach helps to focus on the factors that have the greatest impact on the grouping of electrocardiogram signals.

[0037] For example, when calculating the similarity matrix among patients within a group, the primary set of background variables is used to measure the proximity between patients. One possible implementation uses Euclidean distance to calculate the differences in age and medical history variables to generate the similarity matrix. Assuming patients A and B have an age difference of 2 years and both have a history of hypertension, their similarity is high, corresponding to a value of 0.9 in the matrix. This distribution data provides a basis for subsequent weight adjustments.

[0038] For example, when adjusting group weights based on similarity distribution data, patient groups with high similarity can be assigned higher weights to highlight their representativeness. For instance, if the mean similarity within a group reaches 0.85, its weight is adjusted from the initial 1.0 to 1.2. This adjusted ECG signal data grouping better reflects the intrinsic correlation between patients.

[0039] For example, when extracting time-domain heart rate variability parameters for weighted groups, key indicators such as the standard deviation of the RR interval are emphasized. If the standard deviation of the RR interval for a certain group is 45 ms, which is lower than the preset threshold of 50 ms, it is marked as a potentially abnormal group. This marking method helps to quickly screen out groups that may have problems.

[0040] For example, when performing frequency domain heart rate variability analysis on groups of potential anomaly markers, attention is paid to parameters such as the low-frequency to high-frequency power ratio. If a group has a ratio of 2.5, exceeding the normal range of 1.0 to 2.0, its abnormal characteristics are further confirmed. This analysis supplements the limitations of time domain analysis from a frequency domain perspective, improving the comprehensiveness of anomaly detection. For example...

[0041] It should be noted that the above steps are closely integrated with the patient's background and electrocardiogram signal characteristics, forming a multi-level analytical framework from background variables to time-domain and frequency-domain parameters. This method can effectively identify potential abnormal groupings, providing a reliable basis for subsequent clinical decisions, while reducing the possibility of misjudgment and improving analytical efficiency.

[0042] S103. Based on the ECG signal data grouping after compensation processing, obtain the relevant fluctuation index of unstable diagnostic accuracy. If the fluctuation index exceeds the preset threshold, apply a filtering algorithm to the grouping to adjust the boundary, and obtain ECG signal data grouping after boundary adjustment, wherein the grouping after boundary adjustment improves stability.

[0043] By processing data groups of electrocardiogram (ECG) signals, fluctuation index data of diagnostic accuracy within each group is obtained, and the specific distribution of the fluctuation index is determined. Based on the distribution of the fluctuation index, it is determined whether it exceeds a preset threshold range. If it does, a filtering algorithm is applied to the group boundaries to obtain boundary-adjusted data groups. For the boundary-adjusted data groups, signal processing features within each group are extracted, and the correlation between these features and data stability is analyzed to determine the stability parameters of the feature distribution. Based on the stability parameters of the feature distribution, the consistency level of the processed data within each group is obtained. If the consistency level is lower than a preset standard, a secondary smoothing process is performed on the signals within the group to obtain data groups with improved consistency. For the data groups with improved consistency, the changing trends of ECG signals between different groups are analyzed to determine whether the changing trends meet preset stationarity conditions. If not, the data between groups is equalized to obtain equalized data groups. Through the equalized data groups, the index analysis results within each group are extracted, and combined with the diagnostic accuracy requirements, the stability evaluation data of the final group is determined. Based on the stability assessment data of the final group, a signal processing report for each group is generated to determine whether the signal processing has achieved the preset stability target, and the final processed ECG signal data group is obtained.

[0044] By processing the data in groups of electrocardiogram (ECG) signals, we can first obtain the fluctuation index data of diagnostic accuracy within each group, thereby determining the specific distribution of the fluctuation index. For example...

[0045] In one possible implementation, diagnostic accuracy values ​​are calculated for multiple groups separately. Suppose that the accuracy of one group fluctuates between 85% and 92%, while that of another group fluctuates between 78% and 95%. By statistically analyzing the distribution of these fluctuation indicators, it can be clearly seen that some groups have larger fluctuation ranges.

[0046] Specifically, if the preset threshold range is fluctuation of no more than 8%, then groups that exceed this range need to be further processed to avoid instability in the diagnostic results.

[0047] In one possible implementation, when the fluctuation index is determined to exceed a preset threshold range, a filtering algorithm is applied to the data group boundaries to process the data groups and obtain data groups with adjusted boundaries.

[0048] For example, a low-pass filtering algorithm is used to smooth the signal transition region at the boundary. Assuming that a sudden signal change at the original boundary causes a 12% fluctuation in accuracy, the filtered change amplitude is reduced to 5%, resulting in smoother groupings after boundary adjustment and a more concentrated distribution of diagnostic accuracy. This process helps reduce the interference of boundary noise on group stability and improves the overall data reliability. For the boundary-adjusted data groups, signal processing features within each group are extracted, and the correlation between these features and data stability is analyzed to determine the stability parameters of the feature distribution.

[0049] For example, features such as P-wave amplitude and QRS complex duration are extracted, and their standard deviations within the groups are observed. It is assumed that a standard deviation of 0.15 mV for P-wave amplitude corresponds to higher stability, while a deviation exceeding 0.25 mV indicates lower stability. Through this correlation analysis, the contribution of features to stability can be quantified, providing a basis for subsequent optimization.

[0050] Specifically, the consistency level of the signal after processing within each group is obtained based on the stability parameter of the feature distribution. If the consistency level is lower than the preset standard, the signal within the group is subjected to secondary smoothing processing to obtain a data group with improved consistency.

[0051] For example, if the preset consistency standard is 85%, and a group only achieves 72%, then a moving average smoothing is applied, raising the consistency to 88% and resulting in a more uniform signal waveform. This secondary smoothing effectively reduces internal noise and ensures high consistency of data within groups, which is beneficial for the accuracy of subsequent trend analysis. For the data groups with improved consistency, the changing trends of the ECG signals between different groups are analyzed to determine whether the changing trends meet the preset stationarity conditions. If not, the data between groups is equalized to obtain equalized data groups.

[0052] For example, when observing the trend of RR intervals between adjacent groups, if the mean of one group suddenly increases by 15%, which does not meet the stationarity requirement, normalization and equalization are performed to control the difference between groups within 5%. This equalization process can eliminate human bias between groups and enhance the comparability between groups. After equalization, the data are grouped, and the index analysis results within each group are extracted. Combined with the requirements of diagnostic accuracy, the stability assessment data of the final group is determined.

[0053] For example, by comprehensively calculating the coefficient of variation of each group, assuming that the overall coefficient of variation drops to 0.08 after balancing, which meets the diagnostic accuracy requirements, the stability assessment is excellent. This assessment data directly reflects the grouping quality. Based on the stability assessment data of the final grouping, a signal processing report for each group is generated to determine whether the signal processing has achieved the preset stability target, thus obtaining the final processed ECG signal data grouping.

[0054] For example, the report showed that 90% of the grouping stability target was achieved, and the final groupings could be directly used for clinical auxiliary diagnosis. This layered processing approach significantly improved the robustness of ECG signal groupings, reduced the risks associated with diagnostic fluctuations, and enhanced the clinical application value of the analysis.

[0055] S104. Based on the ECG signal data grouping after the boundary adjustment, determine the dynamic requirements set by the judgment criteria. If the grouping shows that the coefficient of variation of heart rate change adjustment is higher than the threshold, generate a dynamic threshold sequence and determine the ECG signal data grouping after the application of the dynamic threshold sequence, wherein the dynamic threshold sequence corresponds to a specific range for an individual.

[0056] By grouping ECG signal data, the variability of heart rate changes in each group is obtained. This variability data is then preliminarily filtered to obtain a set of groups exceeding a preset threshold. For the filtered group set, a dynamic threshold generation method is used to construct a threshold sequence matching the individual range, determining the generated dynamic threshold sequence data. This generated dynamic threshold sequence data is applied to the corresponding data groups to update the ECG signal grouping structure, resulting in a processed signal group set. For the processed signal group set, the matching degree of signals within each group is analyzed. If the matching degree is lower than a preset standard, the signals within the group are smoothed to obtain more consistent group data. Using the more consistent group data, the transition of heart rate changes between groups is obtained. If the transition shows discontinuity, interpolation is performed on the signals between groups to determine the smoothed transition group data. Based on the smoothed transition group data, signal stability parameters within each group are extracted, and it is determined whether the stability parameters meet preset requirements, resulting in the final processed ECG signal group data. For the final processed ECG signal group data, the correspondence between the individual range and the dynamic threshold is analyzed to determine the threshold application result for each group.

[0057] The generation of the dynamic threshold sequence includes the following steps: The dynamic threshold generation is modeled as a Markov decision process using a deep Q-network, where the state space is defined as: in, This is the feature vector of heart rate variability. For the patient's background feature vector, For historical diagnostic record feature vectors; The action space is a threshold adjustment vector: in, For the first The adjustment amount for each threshold; Q-value updates follow the Bellman equation: in, For learning rate, As a discount factor, For instant rewards; Design a reward function that balances accuracy and stability: in, For a moment The diagnostic accuracy The false positive rate, For the threshold variance, , ,a, These are the weighting coefficients; A stable threshold policy is learned using the proximal policy optimization algorithm, with the objective function being: in, For strategy ratio, For the estimation of the advantage function, These are the trimming parameters.

[0058] For example, when processing grouped electrocardiogram (ECG) signal data, obtaining data on the variability of heart rate changes in each group is a crucial first step. The degree of variability can be determined by analyzing the range of heart rate fluctuations over a certain time period, with the aim of identifying groups with abnormally large fluctuations. Suppose one group's heart rate fluctuates between 60 and 85 beats per minute, while another group's ranges between 55 and 95 beats per minute. If a preset threshold is set for a fluctuation range not exceeding 20 beats per minute, the latter will be filtered out as a group exceeding the threshold. This initial screening helps to quickly pinpoint the data range that requires further processing.

[0059] For example, when constructing a threshold sequence that matches the individual range using a dynamic threshold generation method for a selected group set, adjustments can be made based on the individual's historical heart rate data and current state. For instance, a patient with a normally low heart rate fluctuation might have a dynamic threshold set to 15 beats per minute, while for a patient with a higher heart rate fluctuation after exercise, the threshold might be relaxed to 25 beats per minute. This method better adapts to the physiological characteristics of different individuals, ensuring the rationality of the threshold sequence.

[0060] For example, when applying dynamic threshold sequences to update the grouping structure of corresponding data groups, signal distribution can be optimized by re-dividing the grouping boundaries. Suppose a segment of signal in the original group is misclassified due to a sudden change in heart rate; after applying dynamic thresholds, the boundary adjustment makes the signal more consistent with individual characteristics, resulting in a processed set of signal groups. This adjustment effectively reduces the possibility of misclassification.

[0061] For example, when analyzing the matching degree of a processed signal group set, if the matching degree of a signal within a certain group is lower than a preset standard of 80%, smoothing is required. This can be achieved by averaging the signals to make the waveform more continuous, ultimately improving the matching degree to over 85%. This more consistent grouped data lays the foundation for subsequent analysis.

[0062] For example, when analyzing heart rate changes between groups of data with higher consistency, if the mean heart rate between two groups suddenly increases from 70 to 90, indicating discontinuity, interpolation can be used to supplement the intermediate value and make the transition smoother; this process can avoid interference from sudden changes in data analysis.

[0063] For example, when extracting signal stability parameters from grouped data after a smooth transition, the standard deviation of heart rate changes can be considered. If the standard deviation is below 2.5, it indicates that the stability meets the preset requirements; otherwise, further optimization is needed. This parameter judgment provides a basis for the reliability of the final grouped data.

[0064] For example, analyzing the correspondence between individual ranges and dynamic thresholds in the final processed ECG signal group data can check whether the threshold application for each group meets expectations. Assuming a dynamic threshold of 20 beats / minute for a certain group, and the actual fluctuation is controlled within 18 beats / minute, the application result is good. This analysis helps verify the suitability of the processing method.

[0065] S105. From the ECG signal data group after the application of the dynamic threshold sequence, extract the potential pattern of misjudgment of normal fluctuations, and judge the abnormal tendency of the potential pattern by comparing the similarity of the group with the historical benchmark data, so as to obtain the ECG signal data group after the abnormal tendency judgment, wherein the group after the abnormal tendency judgment distinguishes the risk of misjudgment.

[0066] By obtaining relevant information on fluctuation extraction from ECG signal data grouping after applying dynamic threshold sequences, the potential patterns hidden within the fluctuation extraction results are analyzed to determine the basic structure of the potential patterns. Based on the basic structure of the potential patterns, a pre-established classification model, specifically using the support vector machine method, is used to initially classify the potential patterns, resulting in a classified pattern set. The classified pattern set is then compared with historical benchmark data to analyze the matching degree. If the matching degree is lower than a preset threshold, it is marked as having an abnormal tendency, resulting in a marked pattern set. From the marked pattern set, a subset of patterns with abnormal tendencies is extracted, and these patterns are then further processed. The system analyzes the distribution of abnormal trends within signal groups to determine their specific locations. Based on these locations, it adjusts the classification boundaries of the signal groups and recalculates the fluctuation characteristics within each group to obtain updated signal groups. The updated signal groups are then analyzed to assess the distribution of misjudgment risk. If the misjudgment risk in a group exceeds a preset standard, the signal data within that group is weighted to determine the final signal groups that differentiate misjudgment risk. Finally, for each final signal group that differentiates misjudgment risk, the system records relevant information on abnormal trends and fluctuations, generating a structured grouping archive for subsequent analysis. The specific calculations for extracting potential patterns of misjudgment and determining abnormal tendencies are as follows: (1) Use variational autoencoders to learn the latent distribution of a large amount of normal ECG data and construct a normal pattern benchmark; (2) For the new data after dynamic thresholding, the reconstruction error and potential spatial deviation are calculated using the VAE: in, For the input sample, For the reconstructed ECG signal, For reconstruction error, For balancing parameters, The KL divergence between the encoded distribution and the standard normal prior; (9) Introducing generative adversarial networks for data augmentation and anomaly detection: Where G is the generator, which generates the ECG signal from random noise, and D is the discriminator, which determines whether the input is real data or generated data. The value function is that the discriminator D attempts to maximize it, while the generator G attempts to minimize it. The expected log probability of the discriminator on real ECG data. The expected log probability of the discriminator on the generated data. For the distribution of real ECG data, Noise distribution as input to the generator For example, when extracting relevant information about fluctuations from ECG signal data groups after applying dynamic threshold sequences, the amplitude and frequency distribution of heart rate fluctuations in each group can be calculated first.

[0067] Specifically, by scanning the variation sequence of RR intervals within a group, short-term fluctuation peaks and long-term trend shifts are extracted, thus forming fluctuation extraction results. This extraction helps to capture the physiological rhythm changes implicit in the signal.

[0068] For example, when analyzing the potential patterns hidden within the results of fluctuation extraction, it is necessary to pay attention to the repeating patterns and periodic characteristics of the fluctuation curve.

[0069] In one embodiment, if a low-frequency oscillation pattern repeating every 5 minutes is found within a certain group, it can be preliminarily identified as a potential respiratory-related pattern. This analysis process can reveal the possible autonomic nervous system regulation mechanisms behind the signal.

[0070] For example, based on the basic structure of the latent pattern, a pre-established classification model is used, specifically the support vector machine method, to perform preliminary classification of the latent pattern. The support vector machine classifies the pattern into categories such as normal rhythmic, accelerated, and irregular patterns through high-dimensional feature mapping.

[0071] Preferably, the input features include fluctuation amplitude, period length, and symmetry. After classification, a clear set of patterns is obtained, which facilitates subsequent targeted processing.

[0072] For example, when comparing a categorized set of patterns with historical baseline data, the current pattern parameters can be compared with a baseline database of patients' past stable periods. If the match is less than 85%, it is marked as an abnormal tendency. This comparison can detect abnormal signals that deviate from the individual's baseline early, improving the personalized accuracy of the test.

[0073] For example, after extracting a subset of patterns with anomaly tendencies from the labeled pattern set, when analyzing their distribution in signal groups, it is possible to statistically determine the time period proportion and concentration location of abnormal patterns.

[0074] Specifically, if abnormal tendencies are concentrated primarily in the nighttime group, then the specific group located during the sleep stage is identified. This distribution analysis helps to pinpoint potential nocturnal cardiac rhythm events.

[0075] For example, when adjusting the classification boundaries of signal groups based on the specific location of anomalies within a group, the boundaries can be expanded or contracted towards the anomaly concentration area, allowing the anomaly pattern to be more completely classified into a single group. After adjustment, the fluctuation characteristics within each group, such as standard deviation and peak density, are recalculated to obtain updated signal groups. This adjustment significantly reduces the risk of cross-group misclassification.

[0076] For example, when analyzing the distribution of misjudgment risk within each updated signal group, if the misjudgment risk of a certain group is higher than 10%, a higher weight is applied to the signals of that group to highlight the abnormal features, ultimately determining the signal groups that distinguish the misjudgment risk. This weighted processing effectively improves the sensitivity of identifying high-risk areas.

[0077] For example, when recording abnormal tendencies and fluctuations within each group to extract relevant information for signal grouping that ultimately distinguishes the risk of misjudgment, a structured profile containing pattern type, distribution location, and risk score can be generated. This profile provides a reliable data foundation for subsequent long-term monitoring and clinical review, supporting more accurate individualized electrocardiogram analysis.

[0078] S106. For the ECG signal data grouping after the abnormal tendency judgment, analyze the intersection point of the abnormal state boundary. If the intersection point is located within the preset boundary interval, fuse the feature vector within the boundary interval to determine the ECG signal data grouping after fusing the feature vector, wherein the grouping after fusing the feature vector enhances the boundary clarity.

[0079] In step S6, the fusion of feature vectors within the boundary interval to enhance boundary clarity includes: ECG feature vectors near the boundary intervals are constructed into a feature graph, where nodes represent feature vectors and edges represent the similarity or temporal correlation strength between features. A graph attention network is applied to aggregate information from neighboring nodes and update the feature representation of each node, thereby allowing features of the same type of abnormality to cluster in space and features of different categories to separate from each other. Spectral clustering analysis is performed on the updated node features to redefine the classification boundaries of abnormal states based on the distribution structure of the data itself. The final grouping with improved boundary clarity is output, where each group corresponds to a cardiac rhythm or myocardial ischemia state with a clearly defined feature. The feature vectors within the fusion boundary interval are calculated as follows: Applying graph attention networks for feature fusion and propagation: in, For nodes In the Layer feature representation, It is a non-linear activation function. For nodes The set of neighboring nodes, The attention coefficient represents the node. For nodes Importance weight, For the first The learnable weight matrix of the layer is used to linearly transform node features. For nodes In the Layer feature representation.

[0080] For example, in the grouping and processing of electrocardiogram signal data after abnormal tendency judgment, the fluctuation range of signal data within each group is first obtained.

[0081] Specifically, by calculating the difference between the maximum and minimum values ​​of the RR interval sequence, the fluctuation range is determined to be 0.15 seconds within a certain group. This range reflects the amplitude of heart rate variability and helps in subsequent comparison with the limit interval.

[0082] For example, when analyzing the overlap between the fluctuation range and a preset limit interval, the normal limit interval can be set to 0.05 to 0.12 seconds. If the overlap area exceeds a preset threshold of 60%, the group is marked as an object to be processed. This marking method can identify potential abnormal groups early, avoid missing high-risk signals, and thus obtain marked signal data groups, improving the targeting of detection.

[0083] In one embodiment, abnormal state-related signal features, such as QRS complex width and ST segment shift, are extracted based on the labeled signal data groups. These features are directly associated with abnormal tendencies and can capture subtle manifestations of arrhythmias.

[0084] Specifically, when analyzing the distribution of extracted signal features within the boundary interval, it can be observed that the feature values ​​mainly fall near the upper boundary of the interval, forming a set of intersections corresponding to the distribution locations. This analysis reveals the interaction points between features and boundaries, facilitating the precise location of abnormal transition regions.

[0085] For example, for a set of intersection locations, the feature vector corresponding to each intersection location is obtained, including magnitude, slope, and duration. A pre-built support vector machine model is used to classify the feature vectors into stable and biased types, resulting in a classified set of feature vectors. This classification improves the structured understanding of the features and supports subsequent clustering decisions.

[0086] In one embodiment, the clustering of each feature vector within a defined boundary range is analyzed using the classified feature vector set. If the clustering density exceeds a preset standard of 0.8 per square unit, the feature vectors within the clustered region are integrated, such as by averaging to form an integrated feature vector group. This integration reduces noise interference and enhances the representativeness of abnormal signals.

[0087] For example, when redefining the boundaries of signal data groups based on the integrated feature vector set, the boundaries can be shifted 0.02 seconds towards the clustering area. This redefinition makes the groups more consistent with the actual anomaly distribution. For the redefined groups, the consistency of signal features within each group is calculated. If the variance is less than 0.01, it is considered highly consistent, resulting in signal data groups with a consistency assessment. This assessment enhances the reliability of the groups.

[0088] Specifically, for the signal data grouped after consistency assessment, the distribution characteristics of abnormal tendencies within each group are obtained, such as the abnormal tendencies being mainly concentrated in the latter half of the group. When analyzing the degree of matching between the distribution characteristics and the boundary intervals, if the matching degree reaches more than 90%, the abnormal state distribution result is confirmed. This final determination helps in timely clinical intervention and significantly improves the accuracy and personalization of ECG monitoring.

[0089] S107. Based on the ECG signal data grouping after the fusion feature vector, evaluate the gradient distribution of severity identification, and obtain the ECG signal data grouping after quantization scale mapping by mapping the grouping to the quantization scale identification, wherein the grouping after quantization scale mapping achieves accurate differentiation of myocardial ischemia features.

[0090] For the ECG signal data grouped after feature vector fusion, the fluctuation characteristics of the signal data within each group are obtained, and the gradient distribution of the fluctuation characteristics in severity is analyzed to obtain the gradient-distributed signal data groupings. Based on the gradient-distributed signal data groupings, signal feature points related to myocardial ischemia are extracted within each group. A pre-established support vector machine model is used to classify the feature points, determining the set of classified signal feature points. For the set of classified signal feature points, the distribution position of the feature points on the quantization scale is analyzed. If the distribution position deviates from the preset threshold range, the deviating feature points are reclassified to obtain the reclassified and adjusted signal feature point set. Through the reclassified and adjusted signal feature point set, the correlation between each feature point and myocardial ischemia identification is obtained. The correlation strength is analyzed, and its distribution pattern on the quantization scale is determined to establish a set of signal feature points after the correlation strength distribution. Based on this set, the signal data within each group is scale-transformed to generate scale-transformed ECG signal data groups. For each scale-transformed ECG signal data group, the degree of matching between the signal features within each group and myocardial ischemia is analyzed. If the degree of matching is lower than a preset threshold, the signal features within the group are redistributed to obtain the final adjusted ECG signal data groups. Finally, the consistency of the signal feature distribution within each group is obtained through the final adjusted ECG signal data groups. Whether the consistency meets a preset standard is determined, and a final ECG signal data group that meets the standard is generated. The gradient distribution for assessing severity identification specifically includes: (10) Construct a multi-task learning model with a shared-private architecture, wherein the shared layer is: in, To input ECG signals or features, For parameters of the shared layer, For shared layer mapping functions, For shared feature representation; (2) The task-specific layer output is: in, The task index includes the degree of ischemia, anomaly type, and risk level. For the first Private parameters for each task; For the first Mapping function for each task; For the first The predicted output for each task; (3) Employ an uncertainty-weighted multi-task loss function: Where K is the total number of tasks; For the first Uncertainty parameters for each task; For the first The loss function for each task; Assign uncertainty weights to automatically adjust the contribution of each task to the total loss; For regularization terms; (11) Apply gradient surgery to prevent gradient conflicts between tasks: in, For the first The gradient of each task with respect to shared parameters; Let be the gradient of the j-th task with respect to the shared parameters; gradient The square of the L2 norm; gradient and The inner product; For indicator functions; For shared parameter gradients; (5) The quantitative mapping function for the severity of myocardial ischemia is: in, For input ECG signal; For feature extraction functions; This is the weight vector; For bias terms; It is a sigmoid function; This represents the percentage of severity of myocardial ischemia.

[0091] For the ECG signal data grouped after feature vector fusion, the fluctuation characteristics of the signal data within each group are obtained, and the gradient distribution of the fluctuation characteristics in terms of severity is analyzed to obtain the signal data grouped after gradient distribution. For example.

[0092] In one embodiment, the rate of change of ST segment offset within each group is first calculated as a fluctuation characteristic to reflect the dynamic evolution of myocardial ischemia.

[0093] Specifically, if the fluctuation characteristics of a certain group gradually increase from 0.1mV to 0.4mV, a gradient distribution is formed. The steeper gradient corresponds to potentially severe ischemic areas. This analysis helps to distinguish between mild and severe risk at an early stage, obtaining more refined signal data groupings after gradient distribution. Based on the signal data groupings after gradient distribution, signal feature points related to myocardial ischemia are extracted within each group. A pre-established support vector machine model is used to classify the feature points, determining the set of classified signal feature points.

[0094] In one embodiment, the extracted feature points include T-wave inversion depth and Q-wave location, which are directly related to myocardial ischemia. Using a support vector machine (SVM) model, the feature points can be classified into ischemic and non-ischemic types. For example, a feature point with a T-wave inversion of 0.3 mV is classified as ischemic. This classification improves the accuracy of structured feature point recognition and determines a more reliable set of classified signal feature points. For the classified set of signal feature points, the distribution of the feature points on the quantization scale is analyzed. If the distribution deviates from a preset threshold range, the deviating feature points are reclassified to obtain a reclassified and adjusted set of signal feature points.

[0095] For example, by setting a quantization scale threshold of 0.2mV to 0.35mV, if a feature point is distributed at 0.45mV, it is reclassified into the high-risk group. This adjustment reduces misjudgments, enhances the accuracy of distribution location, and the resulting adjusted set of signal feature points is more helpful for subsequent correlation analysis. Using the adjusted set of signal feature points, the correlation strength between each feature point and myocardial ischemia identification is obtained, the distribution pattern of the correlation strength on the quantization scale is analyzed, and the set of signal feature points after the correlation strength distribution is determined.

[0096] Specifically, the correlation strength can be calculated by comparing the similarity between feature points and a standard ischemic template. If the strength values ​​are mainly concentrated above 0.85, the distribution pattern shows a strong correlation cluster. This analysis reveals the regularity of ischemic patterns, and the set of signal feature points after determining the correlation strength distribution can better support accurate diagnosis. Based on the set of signal feature points after the correlation strength distribution, scaling processing is performed on the signal data within each group to generate scale-scaled ECG signal data groups.

[0097] In one embodiment, scaling can unify the amplitude to a standardized range, such as scaling the original signal to the 0-1 interval, facilitating cross-group comparisons. This scaling improves signal comparability, and the scaled ECG signal data groups are more suitable for multi-lead fusion analysis. For the scaled ECG signal data groups, the degree of matching between the signal features within each group and myocardial ischemia is analyzed. If the degree of matching is lower than a preset threshold, the signal features within the group are redistributed to obtain the final adjusted ECG signal data groups.

[0098] For example, if the preset threshold is 80%, and a group only matches 65%, the low-matching features are reassigned to adjacent groups. This reassignment optimizes the group boundaries, resulting in final adjusted ECG signal data groups with higher clinical consistency. Using these final adjusted ECG signal data groups, the distribution consistency of signal features within each group is obtained, and it is determined whether the consistency meets preset standards, generating final ECG signal data groups that meet the standards.

[0099] The above embodiments are merely one of the preferred embodiments of the present invention and should not be used to limit the scope of protection of the present invention. Any modifications or refinements made to the main design concept and spirit of the present invention that are not of substantial significance, but solve the same technical problem as the present invention, should be included within the scope of protection of the present invention.

Claims

1. A method for high-frequency ECG feature analysis of myocardial ischemia and complex heart rate abnormalities, characterized in that, Specifically, the steps include the following: S1. Acquire electrocardiogram (ECG) signal data, extract relevant raw records containing high-frequency feature analysis from a pre-established patient database, perform preliminary classification of the raw records using a support vector machine, and obtain pre-classified ECG signal data groups. S2. For the electrocardiogram signal data grouping after the preliminary classification, compensation processing is performed to ignore individual differences. The compensation-processed electrocardiogram signal data grouping is determined by calculating the weight distribution of background variables of each patient in the group. S3. Based on the ECG signal data grouping after compensation processing, obtain the relevant fluctuation index of unstable diagnostic accuracy. If the fluctuation index exceeds the preset threshold, apply the filtering algorithm to the grouping to adjust the boundary and obtain the ECG signal data grouping after boundary adjustment. S4. Based on the ECG signal data grouping after the boundary adjustment, determine the dynamic requirements set by the judgment criteria. If the coefficient of variation of the heart rate change adjustment shown in the grouping is higher than the threshold, generate a dynamic threshold sequence and determine the ECG signal data grouping after the application of the dynamic threshold sequence. S5. Extract potential patterns of misjudgment of normal fluctuations from the ECG signal data group after the application of the dynamic threshold sequence, and judge the abnormal tendency of the potential patterns by comparing the similarity between the group and the historical benchmark data, and obtain the ECG signal data group after the abnormal tendency judgment. S6. For the ECG signal data grouping after the abnormal tendency judgment, analyze the intersection of the abnormal state boundary. If the intersection is located within the preset boundary interval, fuse the feature vector within the boundary interval to determine the ECG signal data grouping after fusing the feature vector. S7. Based on the grouping of ECG signal data after fusion feature vectors, evaluate the gradient distribution of severity identification, and obtain the ECG signal data grouping after quantization scale mapping by mapping the grouping to the quantization scale identification of myocardial ischemia.

2. The ECG high-frequency feature analysis method for myocardial ischemia and complex heart rate abnormalities according to claim 1, characterized in that, Step S1 specifically includes the following sub-steps: acquiring electrocardiogram (ECG) signal data and extracting original records from a pre-established patient database; extracting high-frequency features based on the original records to obtain high-frequency feature data; classifying the high-frequency feature data using a support vector machine to obtain classified ECG signal data groups; calculating heart rate variability indices for the classified ECG signal data groups to determine the initial distribution of heart rate abnormalities. If the heart rate variability index exceeds a preset threshold, the corresponding group is marked as an abnormal group, and the abnormally marked electrocardiogram (ECG) signal data is obtained. Cluster analysis is performed on the abnormally marked ECG signal data to determine subgroups within the abnormal group. Heart rate abnormality patterns are extracted from the subgroups to obtain heart rate abnormality pattern data. The specific calculations for obtaining ECG signal data and performing preliminary classification are as follows: (1) A one-dimensional convolutional neural network is used to automatically extract features from the original ECG signal, wherein the feature map output by the convolutional layer has The volume is calculated as follows: ;in, Indicates the first The input ECG signal of each channel, where C is the total number of leads; For the first Layer The convolution kernel pairs with the first... One channel, This represents a one-dimensional convolution operation. For bias terms, It is the ReLU activation function; (2) A multi-head self-attention mechanism is introduced to enhance the attention to key temporal features. The specific calculation of the attention mechanism is as follows: ;in, , , , The feature matrix extracted by CNN, For learnable weight matrix, The feature dimension scaling factor; (3) The deep features are fused with traditional heart rate variability statistical features and then input into the support vector machine classifier. The specific calculation is as follows: ;in, For support vector weights, For category labels, For kernel function, This is a bias term.

3. The ECG high-frequency feature analysis method for myocardial ischemia and complex heart rate abnormalities according to claim 1, characterized in that, Step S2 specifically includes the following sub-steps: ranking the background variable values ​​by feature importance using random forest to determine the main background variable set; calculating the similarity matrix among patients within the group for the main background variable set to obtain similarity distribution data; adjusting the weight allocation of the compensated ECG signal data groupings based on the similarity distribution data to determine the weighted ECG signal data groupings; and extracting time-domain heart rate variability parameters from the weighted ECG signal data groupings to obtain time-domain parameter data. If the time-domain parameter data exceeds the preset threshold range, the corresponding group is marked as a potential abnormal group, and the ECG signal data group marked as potential abnormal is determined. Frequency domain heart rate variability analysis was performed on electrocardiogram signal data groups with potential abnormal markers to obtain frequency domain parameter data; Specifically, compensating for individual differences includes: (1) Apply the model-independent meta-learning framework for individual difference adaptive compensation, where the inner layer update is adapted to specific patients to allow the model to quickly adapt to the uniqueness of individual patients. The specific calculation is as follows: ;in, The base model consists of the learnable parameters of a neural network that has been pre-trained on a large amount of general data. For the first indivual The patient's learning task consists of a small amount of the patient's electrocardiogram data and its corresponding labels; The inner learning rate controls the step size when the model adjusts parameters to adapt to an individual patient; For loss function Regarding the basic model The gradient, the loss in the patient On the task of calculation, To adapt the model parameters, i.e., in the base model Based on the patient The new parameters are obtained after several steps of fast gradient update on the data; (2) The outer layer update achieves the meta-learning objective and is used to optimize the base model. : ;in, To optimize the objective, To distribute tasks from all patients The task sampled from different patients is to sum the results. Use the adapted model In patients The loss calculated on the task; (3) The total loss function of the domain adaptive random forest, used to reduce the differences in data distribution between different patient groups, is as follows: ;in, Total training loss; Cross-entropy loss is used for classification and is used to ensure that the model can accurately identify anomalies. To weigh the parameters, which are used to balance the importance of classification loss and domain loss; Domain loss is used to measure the difference between the source domain (training patient population) and the target domain (new patient feature distribution). Domain loss is measured using the maximum mean difference: ;in, For the first Each source domain is an existing, labeled training database sample; No. The target domain is the newly arrived patient sample that needs to be analyzed; These represent the number of samples in the source domain and the target domain, respectively. The feature mapping function maps the original data to the reproducing kernel Hilbert space; For the regenerating nucleus Hilbert space; The norm in the RKHS space is the distance metric; S25, Generate patient-specific weights based on meta-learning results : ;in, This is a temperature parameter used to control the sharpness of the weight distribution; This is a similarity measurement function; In response to the patient The subsequent personalized model; In response to the patient The subsequent personalized model; Basic meta-model; This refers to the total number of patients.

4. The ECG high-frequency feature analysis method for myocardial ischemia and complex heart rate abnormalities according to claim 1, characterized in that, Step S3 specifically includes the following sub-steps: Processing the ECG signal data into groups to obtain fluctuation index data of diagnostic accuracy within each group, and determining the specific distribution of the fluctuation index; based on the distribution of the fluctuation index, determining whether it exceeds a preset threshold range; if it exceeds the preset threshold range, applying a filtering algorithm to the group boundaries of the data groups to obtain boundary-adjusted data groups; for the boundary-adjusted data groups, extracting signal processing features within each group, analyzing the correlation between features and data stability, and determining the stability parameters of the feature distribution; based on the stability parameters of the feature distribution, obtaining the data consistency level after signal processing within each group; if the consistency level is lower than a preset threshold, the data processing ... If a standard is set, the signals within each group are subjected to secondary smoothing to obtain data groups with improved consistency. For the data groups with improved consistency, the trend of ECG signals between different groups is analyzed to determine whether the trend meets the preset stability conditions. If not, the data between groups is equalized to obtain equalized data groups. Through the equalized data groups, the index analysis results within each group are extracted, and combined with the requirements of diagnostic accuracy, the stability assessment data of the final group is determined. Based on the stability assessment data of the final group, a signal processing report for each group is generated to determine whether the signal processing has achieved the preset stability target, thus obtaining the final processed ECG signal data groups.

5. The ECG high-frequency feature analysis method for myocardial ischemia and complex heart rate abnormalities according to claim 1, characterized in that, S4 specifically includes the following sub-steps: by grouping the data of electrocardiogram signals, obtain the variation data of the heart rate change of each group, perform preliminary screening on the variation data, and obtain a set of groups that exceed a preset threshold; For the selected group set, a dynamic threshold generation method is used to construct a threshold sequence matching the individual range, and the generated dynamic threshold sequence data is determined. Based on the generated dynamic threshold sequence data, it is applied to the corresponding data group to update the grouping structure of the ECG signals, resulting in a processed signal group set. For the processed signal group set, the matching degree of signals within each group is analyzed. If the matching degree is lower than a preset standard, the signals within the group are smoothed to obtain more consistent group data. Using the more consistent group data, the transition of heart rate changes between groups is obtained. If the transition shows discontinuity, interpolation is performed on the signals between groups to determine the smoothed transition group data. Based on the smoothed transition group data, signal stability parameters within each group are extracted, and it is determined whether the stability parameters meet preset requirements, resulting in the final processed ECG signal group data. For the final processed ECG signal group data, the correspondence between individual ranges and dynamic thresholds is analyzed to determine the threshold application results for each group; The generation of the dynamic threshold sequence includes the following steps: The dynamic threshold generation is modeled as a Markov decision process using a deep Q-network, where the state space is defined as: ;in, This is the feature vector of heart rate variability. For the patient's background feature vector, For historical diagnostic record feature vectors; The action space is a threshold adjustment vector: ;in, For the first The adjustment amount for each threshold; Q-value updates follow the Bellman equation: ;in, For learning rate, As a discount factor, For instant rewards; Design a reward function that balances accuracy and stability: ;in, For a moment The diagnostic accuracy The false positive rate, For the threshold variance, , ,a, These are the weighting coefficients; A stable threshold policy is learned using the proximal policy optimization algorithm, with the objective function being: ;in, For strategy ratio, For the estimation of the advantage function, These are the trimming parameters.

6. The ECG high-frequency feature analysis method for myocardial ischemia and complex heart rate abnormalities according to claim 1, characterized in that, Step S5 specifically includes the following sub-steps: Obtaining relevant information on fluctuation extraction from the ECG signal data grouping after applying the dynamic threshold sequence; analyzing the potential hidden patterns in the fluctuation extraction results to determine the basic structure of the potential patterns; based on the basic structure of the potential patterns, using a pre-established classification model, specifically the support vector machine method, to perform preliminary classification of the potential patterns, obtaining a classified pattern set; comparing the classified pattern set with historical benchmark data to analyze the matching situation between the pattern set and the historical benchmark data; if the matching degree is lower than a preset threshold, marking it as having an abnormal tendency, obtaining a marked pattern set; extracting a subset of patterns with abnormal tendencies from the marked pattern set. For the aforementioned pattern subset, its distribution within the signal groups is analyzed to determine the specific location of the abnormal tendency within the groups. Based on the specific location of the abnormal tendency within the groups, the classification boundaries of the signal groups are adjusted. For the adjusted groups, the fluctuation characteristics within each group are recalculated to obtain updated signal groups. Through the updated signal groups, the distribution of misjudgment risk within each group is analyzed. If the misjudgment risk within a certain group is higher than a preset standard, the signal data within that group is weighted to determine the final signal groups that distinguish misjudgment risk. For the final signal groups that distinguish misjudgment risk, the relevant information on the abnormal tendency and fluctuation extraction within each group is recorded to generate a structured group file, resulting in signal data groups that can be used for subsequent analysis. The specific calculations for extracting potential patterns of misjudgment and determining abnormal tendencies are as follows: (1) Use variational autoencoders to learn the latent distribution of a large amount of normal ECG data and construct a normal pattern benchmark; (2) For the new data after dynamic thresholding, the reconstruction error and potential spatial deviation are calculated using the VAE: ;in, For the input sample, For the reconstructed ECG signal, For reconstruction error, For balancing parameters, The KL divergence between the encoded distribution and the standard normal prior; (4) Introduce generative adversarial networks for data augmentation and anomaly detection: Where G is the generator, which generates the ECG signal from random noise, and D is the discriminator, which determines whether the input is real data or generated data. The value function is that the discriminator D attempts to maximize it, while the generator G attempts to minimize it. The expected log probability of the discriminator on real ECG data. The expected log probability of the discriminator on the generated data. For the distribution of real ECG data, The noise distribution is the input to the generator.

7. The ECG high-frequency feature analysis method for myocardial ischemia and complex heart rate abnormalities according to claim 1, characterized in that, In step S6, the fusion of feature vectors within the boundary interval to enhance boundary clarity includes: ECG feature vectors near the boundary intervals are constructed into a feature graph, where nodes represent feature vectors and edges represent the similarity or temporal correlation strength between features. A graph attention network is applied to aggregate information from neighboring nodes and update the feature representation of each node, thereby allowing features of the same type of abnormality to cluster in space and features of different categories to separate from each other. Spectral clustering analysis is performed on the updated node features to redefine the classification boundaries of abnormal states based on the distribution structure of the data itself. The final grouping with improved boundary clarity is output, where each group corresponds to a cardiac rhythm or myocardial ischemia state with a clearly defined feature. The feature vectors within the fusion boundary interval are calculated as follows: Applying graph attention networks for feature fusion and propagation: ;in, For nodes In the Layer feature representation, It is a non-linear activation function. For nodes The set of neighboring nodes, The attention coefficient represents the node. For nodes Importance weight, For the first The learnable weight matrix of the layer is used to linearly transform node features. For nodes In the Layer feature representation.

8. The ECG high-frequency feature analysis method for myocardial ischemia and complex heart rate abnormalities according to claim 1, characterized in that, Step S7 specifically includes the following sub-steps: For the ECG signal data grouping after fusion of feature vectors, obtain the signal data fluctuation characteristics in each group, analyze the gradient distribution of the fluctuation characteristics in severity, and obtain the signal data grouping after gradient distribution; According to the signal data grouping after gradient distribution, extract signal feature points related to myocardial ischemia in each group, use a pre-established support vector machine model to classify the feature points, and determine the set of classified signal feature points. For the classified set of signal feature points, analyze the distribution of feature points on the quantization scale. If the distribution deviates from the preset threshold range, the deviated feature points are reclassified to obtain the reclassified and adjusted set of signal feature points. By classifying and adjusting the set of signal feature points, the correlation strength between each feature point and the identification of myocardial ischemia is obtained, the distribution pattern of the correlation strength on the quantization scale is analyzed, and the set of signal feature points after the correlation strength distribution is determined. Based on the set of signal feature points after correlation strength distribution, scaling is performed on the signal data within each group to generate scaled ECG signal data groups. For the scaled ECG signal data groups, the matching degree between the signal features within each group and myocardial ischemia is analyzed. If the matching degree is lower than a preset threshold, the signal features within the group are redistributed to obtain the final adjusted ECG signal data groups. Through the final adjusted ECG signal data groups, the distribution consistency of signal features within each group is obtained, and it is determined whether the consistency meets the preset standard to generate the final ECG signal data groups that meet the standard. The gradient distribution for assessing severity identification specifically includes: (5) Construct a multi-task learning model with a shared-private architecture, wherein the shared layer is: ;in, To input ECG signals or features, For parameters of the shared layer, For shared layer mapping functions, For shared feature representation; (2) The task-specific layer output is: ; in, The task index includes the degree of ischemia, anomaly type, and risk level. For the first Private parameters for each task; For the first Mapping function for each task; For the first The predicted output for each task; (3) Employ an uncertainty-weighted multi-task loss function: Where K is the total number of tasks; For the first Uncertainty parameters for each task; For the first The loss function for each task; Assign uncertainty weights to automatically adjust the contribution of each task to the total loss; For regularization terms; (6) Apply gradient surgery to prevent gradient conflicts between tasks: ;in, For the first The gradient of each task with respect to shared parameters; Let be the gradient of the j-th task with respect to the shared parameters; gradient The square of the L2 norm; gradient and The inner product; For indicator functions; For shared parameter gradients; (5) The quantitative mapping function for the severity of myocardial ischemia is: ;in, For input ECG signal; For feature extraction functions; This is the weight vector; For bias terms; It is a sigmoid function; This represents the percentage of severity of myocardial ischemia.

9. A system based on the ECG high-frequency feature analysis method for myocardial ischemia and complex heart rate abnormalities according to any one of claims 1 to 8, characterized in that, include: The data acquisition module is used to acquire electrocardiogram signal data and extract raw records from the patient database; The preliminary classification module is used to perform preliminary classification of the original records using a support vector machine to obtain preliminarily classified electrocardiogram signal data groups; The individual difference compensation module is used to compensate for the neglect of individual differences using random forest and to determine the grouping after compensation. The boundary adjustment module is used to apply a filtering algorithm to the grouping based on the fluctuation index to adjust the boundary, and obtain the grouping after boundary adjustment. The dynamic threshold generation module is used to generate dynamic threshold sequences and apply them to grouping. The misjudgment pattern recognition module is used to extract potential patterns of misjudgment of normal fluctuations and to determine abnormal tendencies; The feature fusion module is used to fuse feature vectors within the boundary interval to enhance the clarity of the boundary. The severity quantization module is used to evaluate the gradient distribution of severity identification and map it to a quantization scale.

Images