Coronary microvascular dysfunction detection system based on slow feature analysis

Abstract

The application discloses a coronary microvascular dysfunction detection system based on slow feature analysis, relates to the technical field of biomedical signal processing and computer-aided diagnosis, and comprises the following modules: a data acquisition module, which acquires multi-stage electrocardiosignal of a to-be-detected person; a pretreatment module, which performs pretreatment on the multi-stage electrocardiosignal to obtain pretreated multi-stage electrocardiosignal; a model construction module, which constructs a slow feature analysis model, trains the slow feature analysis model, and obtains a trained slow feature analysis model; a slow feature extraction module, which inputs the pretreated multi-stage electrocardiosignal into the slow feature analysis model to perform feature extraction, obtains multiple slow features, and obtains core diagnostic features after screening; and a classification detection module, which inputs the core diagnostic features into a pre-trained classifier to perform diagnosis and obtain a coronary microvascular dysfunction diagnosis result. By introducing slow feature analysis and combining rest-load-recovery multi-stage information, the detection precision of coronary microvascular dysfunction is significantly improved.

Description

Technical Field

[0001] This invention relates to the fields of biomedical signal processing and computer-aided diagnostic technology, and in particular to a coronary microvascular dysfunction detection system based on slow feature analysis. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] Coronary microvascular dysfunction (CMVD) is a common cardiovascular disease characterized by structural or functional abnormalities in the coronary microcirculation, leading to insufficient blood supply to the myocardium. Epidemiological studies have shown that CMVD has a higher incidence in patients with angina symptoms but normal coronary angiography and is closely associated with adverse cardiovascular events.

[0004] Currently, the clinical diagnosis of CMVD faces significant challenges. Conventional coronary angiography cannot detect microvascular lesions, and traditional electrocardiogram (ECG) examinations have low sensitivity for CMVD, primarily due to the lack of specific morphological changes. Existing ECG analysis methods mostly focus on macroscopic waveforms (such as ST segment shift and T wave inversion) or time-frequency domain statistical features, making it difficult to capture the subtle, dynamic, and nonlinear electrophysiological changes caused by CMVD. ECG signals are inherently non-stationary and nonlinear bioelectrical signals; CMVD can disrupt the spatiotemporal synchronicity of myocardial repolarization, causing local electrical activity disturbances, manifesting as multi-scale, multi-lead, and weak signal variations. Furthermore, the heart's ability to regulate load changes (such as from resting to exercise load and then to recovery) is a crucial window for assessing microcirculatory functional reserve; however, existing research is largely limited to single resting state analysis, neglecting the rich pathophysiological information contained in the dynamic evolution of the "resting-load-recovery" three-stage process. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, this invention provides a coronary microvascular dysfunction detection system based on slow feature analysis, which can extract deep, slowly changing electrophysiological features from multi-stage electrocardiogram signals, achieving accurate and non-invasive detection of CMVD to assist doctors in diagnosis.

[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: In a first aspect, the present invention provides a coronary artery microvascular dysfunction detection system based on slow feature analysis, comprising: The data acquisition module is used to acquire multi-stage electrocardiogram signals from the subject. The preprocessing module is used to preprocess multi-stage ECG signals to obtain preprocessed multi-stage ECG signals; The model building module is used to build a slow feature analysis model and train it to obtain a trained slow feature analysis model. The slow feature extraction module is used to input the preprocessed multi-stage ECG signal into the slow feature analysis model for feature extraction, resulting in multiple slow features. After screening, the core diagnostic features are obtained. The classification and detection module is used to input the core diagnostic features into a pre-trained classifier for diagnosis, and obtain the diagnostic results of coronary microvascular dysfunction.

[0007] A further technical solution is that the multi-stage electrocardiogram signal is a 12-lead electrocardiogram signal of the subject in three stages: resting state, overload state, and recovery state.

[0008] Further technical solutions involve preprocessing multi-stage electrocardiogram signals, specifically including: Multi-stage ECG signals are resampled to a uniform length to obtain multi-stage ECG signals with consistent length. Filter and denoise the multi-stage ECG signals of uniform length to obtain the denoised multi-stage ECG signals; The denoised multi-stage ECG signal was standardized to obtain the standardized multi-stage ECG signal.

[0009] A further technical solution is that the slow feature analysis model performs whitening processing on multi-stage electrocardiogram signals before feature extraction to remove the correlation between features.

[0010] In a further technical solution, the slow feature analysis model sequentially includes an input layer, a preprocessing layer, a temporal convolutional network layer, and an output layer. The temporal convolutional network layer sequentially includes an input layer, a first convolutional layer, a second convolutional layer, an adaptive pooling layer, and a fully connected layer.

[0011] A further technical solution is that the loss function of the slow feature analysis model is expressed as:

[0012] in, For loss function, For slowness loss term, To balance the hyperparameters contributed by each constraint term, For the first The variance of each slow output feature. For the first A slow output feature, For the first A slow output feature, For time indexing, To balance hyperparameters, This represents the total number of segments across the three stages. For the first The linear transformation matrix corresponding to each time segment For the first The linear transformation matrix corresponding to each time segment This represents the Frobenius norm.

[0013] A further technical solution involves calculating the slowness score of each slow feature and selecting a predetermined number of slow features as core diagnostic features based on their slowness scores from high to low.

[0014] In a further technical solution, the slowness score is inversely proportional to the feature change rate.

[0015] In a second aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the contents of the coronary microvascular dysfunction detection system based on slow feature analysis as described in the first aspect.

[0016] Thirdly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the contents of the coronary microvascular dysfunction detection system based on slow feature analysis as described in the first aspect.

[0017] The above one or more technical solutions have the following beneficial effects: This invention introduces slow feature analysis and makes key improvements to the traditional slow feature analysis algorithm. By introducing a dynamic temporal smoothing penalty term, the model is forced to learn features that evolve smoothly at different stages of the electrocardiogram signal (resting-load-recovery), which more accurately depicts the dynamic changes in cardiac functional reserve in CMVD patients and significantly improves the physiological relevance and diagnostic accuracy of the features.

[0018] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0020] Figure 1 This is a flowchart of the coronary microvascular dysfunction detection system based on slow feature analysis according to an embodiment of the present invention. Detailed Implementation

[0021] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0022] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0023] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0024] To address the challenges of detecting coronary microvascular dysfunction (CMVD) and the low sensitivity of traditional electrocardiogram analysis methods in existing technologies, this invention proposes a coronary microvascular dysfunction detection system based on slow feature analysis. This system boasts advantages such as high detection accuracy, strong interpretability, and the ability to reflect the dynamic evolution of the disease.

[0025] Example 1 like Figure 1 As shown, this embodiment discloses a coronary artery microvascular dysfunction detection system based on slow feature analysis, including: The data acquisition module is used to acquire multi-stage electrocardiogram signals from the subject. In this embodiment, 12-lead electrocardiogram signals were acquired from the subject during three phases: resting state, overload state, and recovery state. The overload state refers to the subject being under exercise or drug load, and is used to assess the functional reserve of myocardial microcirculation.

[0026] ECG signals from the resting, overload, and recovery phases were all 10 seconds long, with a sampling rate of 500 Hz. The sliding window construction unit had a window length of 250 sampling points and an overlap rate of 50%, ensuring that the generated time series covered multiple complete cardiac cycles.

[0027] A specific example involves acquiring 12-lead electrocardiogram (ECG) records from 30 patients clinically diagnosed with coronary microvascular dysfunction (CMVD). Each patient's ECG record covered three physiological phases: rest, stress, and recovery. A 10-second segment of the 12-lead ECG signal was acquired for each phase, with a sampling rate of 500Hz, resulting in a total of 5000 sampling points. All data were stored in .mat file format in a designated directory. The data loading module automatically traversed all subject directories, read each .mat file, and attempted to extract the ECG signal matrix from the file.

[0028] The preprocessing module, connected to the data acquisition module, is used to preprocess the multi-stage electrocardiogram (ECG) signals to obtain the preprocessed multi-stage ECG signals. In this embodiment, the preprocessing module preprocesses the acquired electrocardiogram (ECG) signal, including a resampling unit, a filtering unit, and a normalization unit.

[0029] The resampling unit is used to resample all ECG signals to a uniform length, achieving timing alignment. The resampling unit uniformly resamples all ECG signals to a set number of sampling points (e.g., 5000) to ensure consistent time length.

[0030] The filtering unit is used to perform bandpass filtering (e.g., 0.5-40Hz) on the signal to remove noise and baseline drift. The filtering unit uses a 4th-order Butterworth bandpass filter with a passband frequency of 0.5-40Hz to filter the signal in each lead to remove baseline drift and high-frequency noise.

[0031] The normalization unit is used to Z-score normalize the signal from each lead to eliminate amplitude differences. The formula is:

[0032]

[0033] in, The standardized electrocardiogram signal, The original electrocardiogram signal. The mean value of the electrocardiogram signal. The standard deviation of the electrocardiogram signal. It is a very small constant to prevent division by zero; The length of the electrocardiogram (ECG) signal sequence. The first in the electrocardiogram signal sequence The amplitude of each sampling point. The standardized signal has a mean of 0 and a standard deviation of 1, eliminating amplitude differences between different subjects and different leads.

[0034] After the ECG signal is standardized, the physiological stage (resting, stress, recovery) of each ECG sample is assigned according to the keywords in the file name (such as "rest", "stress", "recovery") or by taking the modulo of the numbers in the file name, and the temporal context is constructed.

[0035] The preprocessing module also includes a control group generation unit, which is activated when only CMVD patient data is acquired. This unit is used to generate synthetic normal control group data with a distribution similar to the original data through synthetic enhancement techniques. Specifically, it calculates the mean and standard deviation of ECGs for all CMVD patients, adds Gaussian noise and performs Gaussian smoothing to generate control ECG signals that conform to normal physiological characteristics and have a distribution similar to the original data, thereby ensuring that the sample size of the two groups is balanced across the three stages.

[0036] Furthermore, when generating the control group, the number of normal samples generated in each stage is dynamically adjusted according to the number of CMVD samples in that stage (the target is 60% of the number of CMVD samples, and no less than 10), to ensure that the sample size of the two groups is balanced in the three stages.

[0037] The model building module is used to build a slow feature analysis model and train it to obtain a trained slow feature analysis model. In this embodiment, a Slow Feature Analysis (SFA) model is built and trained using the PyTorch framework. The model structure is as follows: (1) Input layer Receive input sequence That is, the preprocessed multi-stage electrocardiogram signal (multi-stage time series data).

[0038] (2) Pretreatment layer (whitening treatment) Before feature extraction, the input data is whitened (spherized) to remove correlations between features and ensure the stability of model learning. Specifically, the mean and covariance matrices of the input data are calculated, and eigenvalue decomposition is performed to obtain the whitening matrix.

[0039] (3) Temporal convolutional network layer (core network structure) To achieve slow feature analysis, this system constructs a network structure based on one-dimensional temporal convolution to replace the traditional linear transformation matrix, thereby better capturing local temporal dependencies in electrocardiogram signals. The specific network structure is as follows: Input: Preprocessed multi-stage ECG signal, with shape (batch_size, 12, 5000), where 12 is the number of leads and 5000 is the duration.

[0040] The first convolutional layer is a one-dimensional convolutional layer with 12 input channels, 32 output channels, a kernel size of 15, a stride of 1, and padding of 7. It is followed by a batch normalization layer and a ReLU activation function.

[0041] The second convolutional layer is a one-dimensional convolutional layer with 32 input channels, 64 output channels, a kernel size of 15, a stride of 1, and padding of 7. It is followed by a batch normalization layer and a ReLU activation function.

[0042] Adaptive pooling layer: Performs global average pooling on the temporal dimension, compressing the temporal length to 1, and obtaining a feature vector with shape (batch_size, 64).

[0043] Fully connected layer: A linear layer that maps 64-dimensional feature vectors to an output dimension of 8 (i.e., outputs 8 slow features).

[0044] For the input sequence The entire learning process of the network can be formally represented as:

[0045] in, For the first The output slow feature at time 1 The value, This refers to the entire network model (including operations such as convolution, pooling, and fully connected layers). This refers to all the learnable parameters of the network (including convolutional kernel weights, fully connected layer weights, and biases).

[0046] (4) Loss function The goal of model training is to learn the transformation matrix that makes the output features change most slowly. The loss function consists of three parts: slowness loss, variance constraint, and decorrelation constraint. The slowness loss minimizes the mean square value of the time derivative of the output signal; the variance constraint ensures that the variance of the output signal is 1, preventing all output signals from approaching a constant (degeneracy); and the decorrelation constraint forces different output features to be orthogonal to each other, extracting non-redundant information.

[0047] All three of the following conditions must be met: Zero mean constraint:

[0048] Unit variance constraint:

[0049] Remove relevant constraints:

[0050] This system introduces a dynamic temporal smoothing penalty term based on the traditional slow feature analysis loss function to better capture the slow dynamic characteristics that change over time in multi-stage ECG signals. The overall goal of model training is to minimize the following improved loss function. :

[0051] in, This is the slowness loss term, used to minimize the rate at which the output signal changes over time; For the first The variance of each slow output feature; Hyperparameters used to balance the contributions of various constraint terms; For the first A slow-output feature; For the first A slow-output feature; For time indexing; It is a variance constraint term that ensures the output signal has unit variance and avoids degeneracy. It removes relevant constraint terms, forces different output features to be orthogonal to each other, and extracts non-redundant information. For dynamic time-series smoothing penalty term, where, It is a balancing hyperparameter. It is the total number of segments in the "rest-load-recovery" three-stage process. For the first The linear transformation matrix corresponding to each time segment For the first The linear transformation matrix corresponding to each time segment denoted as the Frobenius norm. The dynamic temporal smoothing penalty term forces the ECG features learned by the model to maintain continuity and consistency during pathophysiological state transitions by constraining the smooth evolution of the transformation matrix of adjacent time segments (e.g., from rest to load, from load to recovery), thereby more accurately capturing the slow electrical remodeling process caused by CMVD.

[0052] (5) Output layer After extracting features, the slow feature analysis model outputs multiple slow features.

[0053] Furthermore, the training steps for the slow feature analysis model are as follows: A training dataset consisting of electrocardiogram (ECG) signals from patients with coronary microvascular dysfunction was obtained, and a normal ECG signal with a distribution similar to that of the training dataset was generated using synthetic enhancement techniques as a control group. A complete training dataset containing patients with coronary microvascular dysfunction and a normal control group was constructed. Each segment of ECG signal in the complete training dataset is divided into sliding window segments to generate multiple time series subsequences as training batches. The training batch is input into the slow feature analysis model, and the model parameters are iteratively updated by minimizing the preset loss function until convergence. The loss function includes a slowness loss term, a variance constraint term, and a decorrelation constraint term.

[0054] The preprocessed long-term ECG signal was divided into multiple time subsequences using a sliding window; these subsequences were then input into the model in batches; training was performed using the Adam optimizer and a cosine annealing learning rate scheduler; the model was trained iteratively for 50 epochs, and the average loss for each epoch was recorded. After training, the model's linear transformation matrix W learned the optimal projection direction.

[0055] In a preferred embodiment of the present invention, the sliding window has a length of 250 sampling points, corresponding to 0.5 seconds, with an overlap rate of 50%; the subsequences are input into the model in batches, with batch_size=32; during model training, the learning rate lr=0.01.

[0056] The slow feature extraction module is connected to the preprocessing module and the model building module respectively. It is used to input the preprocessed multi-stage electrocardiogram signal into the slow feature analysis model for feature extraction, obtain multiple slow features, and obtain the core diagnostic features after screening. In this embodiment, the slow feature extraction module receives the trained slow feature analysis model from the model building module, which is used to extract multiple slow features reflecting the steady-state properties of cardiac electrical activity from the preprocessed three-stage ECG signal.

[0057] The preprocessed ECG signal data of the test subject is input into the trained SFA model, and the output features are obtained through forward propagation calculation. Its shape is ( , , ),in, The number of input samples (i.e., the number of time subsequences). The temporal length retained for each subsequence after passing through the model (usually 1 after pooling layers). This refers to the feature dimensions output by the slow feature analysis model. Specifically, the long-term ECG signal (preprocessed ECG signal of the subject) is divided into multiple time subsequences. These multiple time subsequences are input into the slow feature analysis model to extract slow features, resulting in multiple slow features. Core diagnostic features are then selected from these slow features by calculating the slowness score for each feature and selecting a predetermined number of slow features as the core diagnostic feature set, ranked from highest to lowest score.

[0058] The slowness score is expressed as follows:

[0059] in, For the first The mean of temporal fluctuations of the slow-motion characteristic. For the first A slow feature at time... The value, For the first A slow feature at time... The value, The total duration of the signal is denoted as . This score quantifies the average energy of a feature changing over time, and the score is directly proportional to the rate of feature change (i.e., the lower the score, the slower the feature changes). For ease of intuitive understanding, a slowness score is defined in this invention. for:

[0060] in, To prevent extremely small constants from being divided by zero, in the context of this embodiment, the slowness score is inversely proportional to the feature change rate; a higher score indicates a slower feature change. Subsequently, the model sorts the slowness scores from highest to lowest and selects a predetermined number of slow features as the core diagnostic feature set.

[0061] The slow feature analysis model is a linear model with an input dimension of 12, corresponding to 12 leads, and an output dimension of 8, outputting a total of 8 slow features. The core diagnostic feature set includes the top three slow features in terms of slowness score, namely SF6, SF5, and SF4.

[0062] The classification and detection module, connected to the slow feature extraction module, is used to input core diagnostic features into a pre-trained classifier for diagnosis, thereby obtaining diagnostic results for coronary microvascular dysfunction.

[0063] In this embodiment, firstly, based on the core diagnostic feature set selected by the slow feature extraction module (e.g., the top three slowness scores: SF6, SF5, and SF4), all slow features output by the model are extracted. Extract the corresponding features. Then, average these core features over time to obtain a 3D feature vector for each sample. .Will With the corresponding tags The core diagnostic feature dataset consists of CMVD or NORMAL. This dataset is divided into training and test sets in a 7:3 ratio, with stratified sampling to ensure balanced class distribution. Multiple classifiers (such as XGBoost, Random Forest, and Support Vector Machine) are used to train the training set, and cross-validation is employed to select the optimal model, ultimately yielding a well-trained classifier for detecting CMVD.

[0064] The trained classifier is used to distinguish between patients with coronary microvascular dysfunction and those without. Core diagnostic features are input into the trained classifier for diagnosis, yielding a diagnosis of coronary microvascular dysfunction, i.e., whether the subject has coronary microvascular disease.

[0065] The results showed that the slow feature-based classifier could effectively distinguish CMVD patients from normal controls, with an AUC value of over 0.96.

[0066] In summary, the coronary microvascular disease detection system based on multi-stage electrocardiogram slow feature analysis provided by this invention can effectively extract slow features with physiological interpretability. By mining the dynamic information of the three stages of "resting-load-recovery", it achieves high-precision and non-invasive detection of CMVD, which has significant clinical application value and promotion prospects.

[0067] To enhance the interpretability of the model and validate the clinical significance of slow features, feature contribution analysis and multi-stage comparative analysis were conducted.

[0068] The following is an example of analysis: The preprocessed ECG signals of the subjects were input into the trained SFA model, which output multiple slow features. This invention selects the top six slowest-changing features (SF1-SF6) from the output dimensions for subsequent analysis, as these features contain the richest steady-state information of the system. The slowness scores for each feature were calculated, and the results are shown in Table 1. SF6, SF5, and SF4 had the highest slowness scores (62.25, 57.71, and 45.94, respectively) and were selected as the core feature set.

[0069] (1) Slow ordering By calculating and ranking the slowness of the six features across all samples in the test set, a slowness score and usage recommendations were obtained for each feature. A higher slowness score indicates that the feature changes more slowly over time, better reflecting the inherent stability of the system and making it suitable as a core feature for diagnostic models. The ranking results of the importance of slow features are shown in Table 1.

[0070] Table 1 Ranking of the Importance of Slow Features

[0071] The results showed that SF6, SF5, and SF4 had significantly higher slowness scores than other features, and their standard deviations contributed significantly, indicating that these features have high stability and information content in the time-varying process of ECG signals. SF6 had the highest slowness score, reaching 67.42, mainly reflecting the variation of low-frequency oscillation components in the ECG signal; SF5 was second, at 63.18, closely related to the spectral distribution of the ECG signal; SF4 had a score of 49.85, mainly reflecting nonlinear dynamic characteristics. These three features are suitable as core features for classification modeling.

[0072] (2) Correlation analysis To assess the redundancy among features, the correlation coefficient matrix among the features was calculated, and the results are shown in Table 2.

[0073] Table 2 Slow Feature Correlation Matrix

[0074] The results showed that SF1 was strongly negatively correlated with SF5 (-0.81), SF1 was strongly positively correlated with SF6 (0.79), and SF2 was strongly negatively correlated with SF4 (-0.79). This indicates that there is a strong linear relationship among some features.

[0075] (3) Multi-stage comparison The six extracted slow features were grouped into three stages: resting, overload, and recovery. Box plots were then drawn for the CMVD group and the normal control group to visually compare the differences in feature distribution between the two groups at each stage. The slowness scores for each stage are shown in Table 3.

[0076] The extracted slow features were grouped into three stages: resting state, stress state, and recovery state. The variance ratio, stress response change rate, and recovery degree of patients with coronary microvascular dysfunction and normal controls were calculated at each stage to verify the consistency between slow features and the known pathological mechanisms of coronary microvascular dysfunction.

[0077] Table 3 Comparison of Slowness Scores for Slow Features at Each Stage

[0078] The results showed that the slowness scores for all characteristics were highest during the load phase, especially for the three core characteristics SF6, SF5, and SF4, which reached 74.28, 71.62, and 56.19 respectively during the load phase, significantly higher than those during the rest and recovery phases. This phenomenon suggests that under myocardial load, the dynamic changes in electrocardiogram signals are slower and more orderly, possibly reflecting the compensatory regulatory mechanism of the heart in response to load.

[0079] (4) Statistical analysis Independent samples t-tests were performed on each slow trait for both groups at each stage. P-values ​​were calculated to assess the statistical significance of the differences, and the traits with the greatest discriminative potential (e.g., SF1 in the loading stage) were selected based on the size of the P-values. At the same time, the percentage difference in slowness for each trait was calculated. The results are shown in Table 4.

[0080] Table 4 Comparison of slowness scores for each feature in the two groups (all stages combined)

[0081] The results showed that SF6, SF5, and SF4 exhibited the most significant differences in slowness between the two groups. Particularly in SF5, the slowness score was significantly higher in CMVD patients than in non-CMVD patients, suggesting that this feature has good sensitivity to disease status. Furthermore, based on the percentage difference in slowness for each feature, SF5 showed the largest difference, reaching approximately 65.39%, followed by SF4 and SF6 at 52.94% and 27.34%, respectively. This further validated the effectiveness of the first three features in distinguishing between the two groups. Notably, the significantly elevated slowness score of SF5 in CMVD patients indicates that the ECG signals of CMVD patients exhibit slower fluctuations in the frequency domain dimension represented by SF5, which may be related to myocardial ischemia caused by coronary microvascular dysfunction.

[0082] (5) Performance evaluation and classification results of feature combination Based on the slowness score and inter-group differences, it is recommended to use the three slowest features (SF6, SF5, SF4) as the core feature set, SF2 and SF1 as auxiliary features, and SF3 is not considered for inclusion in the model. To verify the classification performance of different feature combinations, this invention feeds the above feature set into an SVM classifier and obtains the final classification results through five-fold cross-validation, as shown in Table 5.

[0083] Table 5 Comparison of classification results for different feature combinations

[0084] The results showed that the optimal combination (SF6, SF5, SF4) achieved the best classification performance, with an accuracy of 92.18%, sensitivity of 92.72%, specificity of 95.34%, and an AUC of 0.96. This result is significantly better than other combinations, indicating that the three slowest features not only have high individual slowness scores, but also have the strongest complementarity, comprehensively reflecting the pathological changes in the electrocardiogram signals of CMVD patients. Adding SF2 slightly decreased classification performance, while adding all features reduced the accuracy to 89.41%. This further validates the importance of feature selection; redundant or noisy features can interfere with the classifier's decision boundary and reduce model performance.

[0085] (6) Generate report It automatically generates a comprehensive assessment report that includes data distribution, feature slowness ranking, significance analysis results, and recommendations for the best diagnostic features, and saves multi-stage comparative visualization charts to provide quantitative evidence for clinical diagnosis.

[0086] The present invention has the following beneficial effects: High-precision detection: By introducing slow feature analysis, the slow electrical reconfiguration process caused by CMVD is captured from a dynamic perspective. Combined with multi-stage information of "resting-load-recovery", the detection accuracy is significantly improved. Experimental results show that the combination of the three slowest features, SF6, SF5, and SF4, achieves an accuracy of 92.18% and an AUC of 0.96 on the SVM classifier, which is far superior to traditional features.

[0087] High physiological interpretability: The features extracted by slow feature analysis have clear physical meaning. The features with the highest slowness (such as SF6, SF5, and SF4) directly correspond to the slowest-changing steady-state components in the electrocardiogram signal. Their differences can be interpreted as pathological mechanisms such as myocardial repolarization synchronicity disruption, abnormal stress response, and delayed recovery in CMVD patients. This provides clinicians with intuitive and understandable diagnostic evidence.

[0088] Non-invasive and convenient: This invention is based on a conventional 12-lead electrocardiogram examination, requiring no invasive procedures or special equipment. It is low in cost, easy to promote in medical institutions at all levels, and can serve as an effective tool for large-scale screening of CMVD.

[0089] Make full use of dynamic information: Unlike traditional single resting state analysis, this invention fully explores cardiac function reserve information by analyzing the dynamic evolution process of the three stages of "rest-load-recovery", and improves the ability to identify dynamic diseases such as CMVD.

[0090] In summary, this invention solves the problem of the difficulty in non-invasively and accurately detecting coronary microvascular lesions using existing technologies. Its key technical solutions include: acquiring 12-lead electrocardiograms (ECGs) from subjects at rest, during stress, and after recovery; constructing and training a slow feature analysis model to extract slowly changing steady-state features from the preprocessed ECG signals and selecting a core feature set based on slowness scores; grouping and comparing the core features according to the three stages to analyze the distribution differences between patients with coronary microvascular lesions and normal controls, and outputting diagnostic results through a classifier. This invention achieves non-invasive detection based on conventional ECGs, the extracted slow features are physiologically interpretable, and multi-stage dynamic analysis significantly improves detection accuracy. It can be used for early screening of coronary microvascular lesions and assists doctors in diagnosis.

[0091] Example 2 The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the contents of the system of Embodiment 1.

[0092] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium storing a computer program that, when executed by a processor, performs the contents of the system of Embodiment 1.

[0093] The contents involved in the apparatus of Embodiments 2 and 3 above correspond to those of System Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0094] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0095] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0096] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A coronary artery microvascular dysfunction detection system based on slow feature analysis, characterized in that, include: The data acquisition module is used to acquire the multi-stage electrocardiogram (ECG) signals of the subject; the multi-stage ECG signals are 12-lead ECG signals of the subject in three stages: resting state, overload state, and recovery state. The preprocessing module is used to preprocess multi-stage ECG signals to obtain preprocessed multi-stage ECG signals; The model building module is used to build a slow feature analysis model and train it to obtain a trained slow feature analysis model. The loss function of the slow feature analysis model introduces a dynamic temporal smoothing penalty term, which smoothly evolves by constraining the transformation matrix of adjacent time segments. The slow feature extraction module is used to input the preprocessed multi-stage ECG signal into the slow feature analysis model for feature extraction, obtaining multiple slow features. A slowness score is calculated for each slow feature, and a predetermined number of slow features are selected as core diagnostic features based on their slowness scores, from high to low. The slowness score is inversely proportional to the rate of feature change. The slowness score is expressed as follows: in, For the first The mean of temporal fluctuations of the slow-motion characteristic. For the first A slow feature at time... The value, For the first A slow feature at time... The value, The total duration of the signal; Define slowness score for: in, To prevent division by zero of extremely small constants; The slow feature analysis model includes, in sequence, an input layer, a preprocessing layer, a temporal convolutional network layer, and an output layer. The temporal convolutional network layer includes, in sequence, an input layer, a first convolutional layer, a second convolutional layer, an adaptive pooling layer, and a fully connected layer. The loss function of the slow feature analysis model is expressed as: in, For loss function, For slowness loss term, To balance the hyperparameters contributed by each constraint term, For the first The variance of each slow output feature. For the first A slow output feature, For the first A slow output feature, For time indexing, To balance hyperparameters, This represents the total number of segments across the three stages. For the first The linear transformation matrix corresponding to each time segment For the first The linear transformation matrix corresponding to each time segment Represents the Frobenius norm; The classification and detection module is used to input the core diagnostic features into a pre-trained classifier for diagnosis, and obtain the diagnostic results of coronary microvascular dysfunction.

2. The coronary artery microvascular dysfunction detection system based on slow feature analysis as described in claim 1, characterized in that, Preprocessing of multi-stage electrocardiogram signals specifically includes: Multi-stage ECG signals are resampled to a uniform length to obtain multi-stage ECG signals with consistent length. Filter and denoise the multi-stage ECG signals of uniform length to obtain the denoised multi-stage ECG signals; The denoised multi-stage ECG signal was standardized to obtain the standardized multi-stage ECG signal.

3. The coronary artery microvascular dysfunction detection system based on slow feature analysis as described in claim 1, characterized in that, The slow feature analysis model performs whitening processing on multi-stage electrocardiogram signals before feature extraction to remove correlations between features.

4. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the contents of the coronary microvascular dysfunction detection system based on slow feature analysis as described in any one of claims 1-3.

5. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the contents of the coronary microvascular dysfunction detection system based on slow feature analysis as described in any one of claims 1-3.

Images