A non-invasive coronary artery disease detection system

Abstract

The application discloses a non-invasive coronary heart disease detection system, comprising: a preprocessing module for filtering out electrocardiosignal noise and extracting relevant information of coronary heart disease from electronic medical records associated with electrocardiosignal; a multi-modal feature extraction module for receiving preprocessed electrocardiosignal and relevant information of coronary heart disease, extracting electrocardiosignal features, time-frequency image features, electrocardiogram vector diagram features and electronic medical record text features; a multi-modal feature fusion module for splicing and fusing electrocardiosignal features, time-frequency image features, electrocardiogram vector diagram features and electronic medical record text features to obtain fusion features; an electrocardio classification module for label prediction of fusion features to obtain detection results. The application establishes the time sequence correlation of electrocardiosignal by using the multi-modal feature extraction module and the multi-modal feature fusion module, and enhances data by using high-level features, so that whether a patient is a coronary heart disease patient can be accurately divided without trauma.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a non-invasive coronary heart disease detection system. Background Technology

[0002] Previously, coronary heart disease (CHD) primarily affected the elderly, but in recent years, it has shown a significant trend of affecting younger people in my country. CHD is mainly caused by factors such as hypertension, unhealthy lifestyle habits, and emotional stress, leading to atherosclerosis of the coronary arteries and subsequently causing myocardial ischemia, hypoxia, and even necrosis. If patients do not receive timely and effective treatment to control the disease, complications such as heart failure and cardiac rupture can occur as the condition progresses, potentially resulting in sudden death and posing a serious threat to their lives and health. Therefore, early clinical diagnosis of CHD is of paramount importance in patient prevention, treatment, and prognosis.

[0003] Currently, coronary angiography is the gold standard for diagnosing heart disease. It primarily determines the presence and severity of coronary artery lesions by observing and analyzing the contrast agent's dispersion within the coronary arteries. While this method boasts high accuracy, it is an invasive procedure, hindering its widespread clinical application. Electrocardiography (ECG), on the other hand, is one of the most common and widely used diagnostic methods in clinical practice. It offers advantages such as being non-invasive, rapid, convenient, and inexpensive. The standard 12-lead ECG remains a fundamental diagnostic tool that profoundly influences the field of cardiology.

[0004] In the field of deep learning, data quality and model selection have a significant impact on the performance of deep learning models. However, in the field of physiological signal classification, such as electrocardiogram (ECG) classification, existing models are still unable to achieve non-invasive early screening for coronary heart disease. Summary of the Invention

[0005] The purpose of this invention is to solve the problem of non-invasive screening of whether an individual is a patient with coronary heart disease, and to provide a non-invasive coronary heart disease detection system.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A non-invasive coronary artery disease detection system, comprising:

[0008] A preprocessing module is used to filter out noise from the electrocardiogram signal and extract relevant information about coronary heart disease from the electronic medical records associated with the electrocardiogram signal;

[0009] The multimodal feature extraction module includes an electrocardiogram (ECG) signal feature extraction unit, a time-frequency image feature extraction unit, an ECG vector graph feature extraction unit, and an electronic medical record text feature extraction unit; it is used to receive the preprocessed ECG signal and related information of the coronary heart disease, and extract ECG signal features, time-frequency image features, ECG vector graph features, and electronic medical record text features;

[0010] The multimodal feature fusion module is used to concatenate and fuse the electrocardiogram signal features, the time-frequency image features, the electrocardiogram vector map features, and the electronic medical record text features to obtain fused features;

[0011] The ECG classification module is used to predict the labels of the fused features to obtain the detection results.

[0012] In some embodiments, the multimodal feature fusion module fuses the electrocardiogram signal features, the time-frequency image features, the electrocardiogram vector map features, and the electronic medical record text features using a self-attention mechanism.

[0013] In some embodiments, the ECG signal feature extraction unit is used to obtain different features from the ECG signal by passing it through at least two feature extractors of different scales, then splicing and fusing the different features, and finally obtaining the ECG signal features through a channel attention mechanism and a residual structure, and finally through an adaptive attention mechanism.

[0014] In some embodiments, the ECG signal feature extraction unit structure is one of the following network structures: convolutional neural network, a combination of CNN and LSTM, a combination of CNN and Transformer, or Transformer.

[0015] In some embodiments, the time-frequency image feature extraction unit is used to perform a short-time Fourier transform on the electrocardiogram signal to generate a two-dimensional image and extract features to obtain the time-frequency image features.

[0016] In some embodiments, the time-frequency image feature extraction unit is used to extract features from the two-dimensional image through a Vit network to obtain the time-frequency image features.

[0017] In some embodiments, the ECG vector map feature extraction unit is used to generate an ECG vector map signal from the ECG signal via Kors regression and then extract the features to obtain the ECG vector map features.

[0018] In some embodiments, the ECG vector map feature extraction unit is used to extract features from the ECG vector map signal through a one-dimensional convolutional network and a Transformer encoder to obtain the ECG vector map features.

[0019] In some embodiments, the test result is one of coronary heart disease, arrhythmia, or good health.

[0020] In some embodiments, the preprocessing module uses at least one of the following methods to filter out ECG signal noise: bandpass filter filtering, least mean square adaptive filtering, wavelet transform-based filtering, statistical model-based filtering method, neural network-based filtering method, maximum-minimum normalization, and sigmoid normalization.

[0021] The present invention has the following beneficial effects:

[0022] The non-invasive coronary heart disease detection system of this invention integrates information from electrocardiograms, spectrograms, electrocardiogram vector maps, and electronic medical records, thereby enhancing the richness of electrocardiogram data. By utilizing a multimodal feature extraction module and a multimodal feature fusion module to establish the temporal correlation of electrocardiogram signals and by using high-level features for data augmentation, it can accurately classify whether a patient has coronary heart disease without invasiveness, providing important guidance for patient prevention, treatment, and prognosis.

[0023] In some embodiments, the following beneficial effects are also present:

[0024] This invention utilizes convolutional kernels with different receptive fields to extract features from different ECG signal segments, and employs attention mechanisms between ECG signals and between modules to extract richer features. Existing technologies mostly use single-lead or 12-lead data, utilizing only partial information from the database. This invention incorporates lead information from the ECG vector map using an ECG vector map feature extraction unit, enabling a more comprehensive observation of cardiac electrical activity and effectively improving accuracy.

[0025] Other beneficial effects of the embodiments of the present invention will be further described below. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the non-invasive coronary heart disease detection system in an embodiment of the present invention;

[0027] Figure 2 This is a schematic diagram of the detection process of the non-invasive coronary heart disease detection system in an embodiment of the present invention;

[0028] Figure 3 This is an example network structure diagram of the electrocardiogram signal feature extraction unit in an embodiment of the present invention;

[0029] Figure 4 This is a network structure diagram of the Vit model instance of the time-frequency image feature extraction unit in this embodiment of the invention;

[0030] Figure 5 This is a flowchart of the training and testing process for the non-invasive coronary heart disease detection system in this embodiment of the invention;

[0031] Figure 6 This is a flowchart illustrating the process of a non-invasive coronary heart disease detection system for screening coronary heart disease in this embodiment of the invention. Detailed Implementation

[0032] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and is not intended to limit the scope and application of the present invention.

[0033] Example

[0034] In the field of physiological signal classification, exemplified by electrocardiogram (ECG) classification, existing models are relatively outdated and often fail to establish temporal dependencies in ECG signals. Furthermore, high-level features of the data are not utilized, making them unsuitable for screening coronary heart disease. This invention establishes temporal correlations in ECG signals and utilizes high-level features for data augmentation to perform early screening for coronary heart disease, enabling accurate screening.

[0035] The non-invasive coronary artery disease detection system of this invention is a multimodal fusion coronary artery disease preliminary screening and arrhythmia detection network (MF-CADNet). This embodiment of the system uses electrocardiogram and electronic medical record data for preliminary coronary artery disease screening, referencing... Figure 1 The non-invasive coronary heart disease detection system of this embodiment includes:

[0036] A preprocessing module is used to filter out noise from the electrocardiogram signal and extract relevant information about coronary heart disease from the electronic medical records associated with the electrocardiogram signal;

[0037] The multimodal feature extraction module includes an electrocardiogram (ECG) signal feature extraction unit, a time-frequency image feature extraction unit, an ECG vector map feature extraction unit, and an electronic medical record text feature extraction unit. It is used to receive the preprocessed ECG signal and related information about coronary heart disease, and to extract ECG signal features using the ECG signal feature extraction unit, time-frequency image features using the time-frequency image feature extraction unit, ECG vector map features using the ECG vector map feature extraction unit, and electronic medical record text features using the electronic medical record text feature extraction unit.

[0038] The multimodal feature fusion module is used to concatenate and fuse the electrocardiogram signal features, the time-frequency image features, the electrocardiogram vector map features, and the electronic medical record text features to obtain fused features;

[0039] The ECG classification module is used to predict the labels of the fused features to obtain the detection results.

[0040] Specifically, such as Figure 2As shown, this embodiment of the invention extracts ECG signal features from the ECG signal, then obtains a time-frequency image by performing short-time Fourier transform and other methods. The time-frequency image features are then extracted using an image feature extractor. Next, the ECG signal is transformed to obtain an ECG vector map, which is then processed by a one-dimensional signal feature extractor to obtain its features. Features such as age and gender are extracted from the electronic medical record and processed by a text encoder to obtain word vector features. The different features are then fused using a cross-module attention mechanism, and finally, a classifier is used for label prediction. In summary, this model integrates information from ECG, spectrogram, vectorcardiogram, and electronic medical record, and utilizes convolutional kernels with different receptive fields to extract features from different signal segments. It also employs inter-lead and inter-module attention mechanisms to extract richer features. This model represents a performance improvement over other existing technologies in arrhythmia detection and provides a new approach for non-invasive screening of coronary heart disease.

[0041] The following is a detailed description of each module in the system of this embodiment:

[0042] I. Preprocessing Module:

[0043] In this embodiment, the preprocessing module is used to acquire electrocardiogram (ECG) signals and electronic medical records, extract relevant information about coronary heart disease from the electronic medical records, and preprocess the ECG signals. Specific relevant information about coronary heart disease includes age, gender, presence of hypertension, and family history of coronary heart disease. The ECG signals are preprocessed using filtering methods to remove noise such as baseline drift and electromyography (EMG) signals. The preprocessing module in this embodiment employs any of the following methods during preprocessing: bandpass filter filtering, LMS (least mean square) adaptive filtering, wavelet transform-based filtering, statistical model-based filtering, neural network-based filtering, max-min normalization, sigmoid (a function) normalization, etc.

[0044] II. Multimodal Feature Extraction Module:

[0045] 2.1 ECG signal feature extraction unit:

[0046] In this embodiment, reference Figure 3 The ECG signal feature extraction unit extracts features from the preprocessed 12-lead ECG signal. The ECG signal feature extraction includes the following steps:

[0047] A1. Feature extraction from ECG signals is performed through multi-branch convolution, and the features from different branches are concatenated. The multi-branch module contains different convolutional kernels in different branches, extracting features through different receptive fields, achieved through dilated convolution.

[0048] A2. The features of multiple channels are refined through SENet (channel attention module) and residual network. SENet is a plug-and-play channel attention module. The feature map output by the convolutional layer has multiple channels. SENet strengthens important features and weakens unimportant features by modeling the correlation between the channels of the feature map.

[0049] As shown in the figure, the instance network architecture for ECG signal feature extraction is shown below.

[0050] A3. The refined features are fed into the Transformer encoder to further extract temporal correlation information. Convolutional operations can only capture local information and cannot establish long-distance connections of global information. The Transformer uses multi-head attention to capture richer feature information, expand the receptive field, and has a stronger global perspective.

[0051] Specifically, the one-dimensional electrocardiogram (ECG) signal is processed by at least two feature extractors of different scales; in this embodiment, four feature extractors of different scales are used. The different features are then concatenated and fused. Next, the signal undergoes channel attention and residual structure processing, and finally, adaptive attention is used to obtain the features of the one-dimensional ECG signal. In other embodiments, a single feature extractor can also be used. After the ECG signal feature extraction unit, the ECG features can be represented as follows:

[0052]

[0053] The network model of the ECG signal feature extraction unit provided in this invention includes any one of the following: convolutional neural network, CNN (a type of convolutional neural network) + LSTM (long short-term memory network), CNN + Transformer, Transformer, etc.

[0054] 2.2 Time-Frequency Image Feature Extraction Unit:

[0055] In this embodiment, the time-frequency image feature extraction unit uses short-time Fourier transform, which can take into account both frequency features and time series variation information. The extraction of time-frequency image features includes the following steps:

[0056] B1. In this embodiment of the invention, frequency domain information within a time window is extracted by windowing the electrocardiogram data and performing a Fourier transform. This embodiment selects the Hanning window as the window function to obtain a time-frequency graph, where the horizontal axis represents time and the vertical axis represents frequency. The spectrum helps us analyze the frequency characteristics of the signal more intuitively.

[0057] For a signal x(t) over a time period, its short-time Fourier transform can be expressed as:

[0058]

[0059] Where f is the frequency, t is the time, τ is the delay, and w(t-τ) is the window function.

[0060] The time-frequency image acquisition method provided in the embodiments of the present invention includes any one of the following: short-time Fourier transform, wavelet transform, etc.

[0061] B2. Use the Vit model (a visual model) for time-frequency image feature extraction. A visual Transformer is used to extract spectrogram features. Vit divides the image into fixed-size patches, projects them into fixed-length vectors, adds special characters (cls) and positional encoding, and feeds them into a transformer encoder. Through a self-attention mechanism, long-range dependencies are established, capturing global information in the sequence. The features encoded by the Vit model can be represented as...

[0062]

[0063] The network model of the time-frequency image feature extraction unit provided in the embodiments of the present invention includes any one of the following: convolutional neural network, CNN+LSTM, CNN+Transformer, Transformer, Vit, etc.

[0064] The Vit model instance network structure used in this embodiment of the invention is as follows: Figure 4 As shown, specifically, the input time-frequency image is uniformly divided into blocks of the same size. Each image block is linearly mapped to obtain the features of each image block. Then, the position encoder embeds the position features of each image block at different positions. Finally, the encoder of the Transformer is used to obtain the feature representation of the time-frequency image.

[0065] 2.3 ECG Vector Graph Feature Extraction Unit:

[0066] VCG records the instantaneous spatial vector of the heart, revealing the sequence of cardiac excitation and changes in instantaneous potentials, while ECG (electrocardiogram) only reflects the action potential on a single lead axis. VCG can more clearly observe the depolarization and repolarization processes of the heart in three dimensions, and can more clearly reflect the physiological electrical activity and pathological state of the heart. ECG, on the other hand, can only record the potential difference of cardiac action potentials on the body surface. The physiological electrical activity and pathological state of the heart can be indirectly inferred from the ECG waveform. VCG is more explicit than ECG in showing the sequence of atrial and ventricular electrical excitation, changes in instantaneous vectors, and spatial location, especially in atrial and ventricular hypertrophy, myocardial infarction, intraventricular conduction block, Wolff-Parkinson-White syndrome, and changes in T vectors, which ECG cannot detect. In this embodiment, the feature extraction steps for the ECG vector map are as follows:

[0067] C1. VCG signal acquisition uses Kors regression (a regression analysis method) to generate an electrocardiogram vector map. The VCG uses the same feature extraction network structure as the electrocardiogram signal. The formula for calculating the VCG signal using Kors regression is as follows:

[0068]

[0069] The methods for obtaining VCG signals provided in this invention include any one of the following: direct measurement method, Kors regression method, etc.

[0070] C2: After obtaining the VCG signal, the network model in the ECG vector graph feature extraction unit is used as the feature extraction network for the VCG signal. The network model includes any one of the following: Convolutional Neural Network, CNN+LSTM, CNN+Transformer, Transformer, Vit, etc. The features output by the network can be represented as...

[0071]

[0072] 2.4 Electronic Medical Record Text Feature Extraction Unit:

[0073] Patients' electronic medical records contain a large amount of personal information. Studies have shown that the risk of cardiovascular disease is related to gender and age. The selected gender and age information is converted into vectors through word embedding and used in neural network training. The features extracted by the network can be represented as...

[0074]

[0075] III. Multimodal Feature Fusion Module:

[0076] The multimodal feature fusion module feeds features from four different aspects into the fusion module for feature fusion. This module fuses output information from time-domain analysis, time-frequency analysis, electrocardiogram vector map analysis, and electronic medical records. It uses a transformer encoder to fuse low-dimensional feature information outputs from multiple modules and relies on attention mechanisms to achieve the fusion and extraction of high-dimensional features. The fused features can be represented as follows:

[0077]

[0078] Finally, the mixed features are processed through a self-attention mechanism to obtain the final features used for classification, which can be represented as follows:

[0079]

[0080] IV. Electrocardiogram Classification Module:

[0081] The ECG classification module is used to predict the labels of the fused features to obtain the detection results.

[0082] The method of using the non-invasive coronary heart disease detection system in this embodiment is as follows:

[0083] Before using the non-invasive coronary heart disease detection system of this embodiment, it is necessary to perform supervised training using the obtained data, including the following steps:

[0084] D1: In supervised classification tasks, data augmentation methods can be used to increase the model's generalization, robustness, and accuracy. These methods include, but are not limited to, time-domain shifting, frequency-domain shifting, adding Gaussian noise, adding pulse waveforms, pruning, and speed adjustment.

[0085] D2: The labeled data is fed into the non-invasive coronary heart disease detection system of this embodiment to obtain the network output, and the loss function is calculated with the label information.

[0086] D3. Set an appropriate learning rate and use the calculated loss to backpropagate and update the parameters of the neural network.

[0087] Specifically, refer to Figure 5 The training and testing process for the non-invasive coronary heart disease detection system in this embodiment is as follows:

[0088] Step 1. Divide the ECG data and electronic medical records into training and testing sets according to the Inter-patient method (feature extraction and classification for different patients);

[0089] Step 2. Feed the training set data into the multimodal model for inference;

[0090] Step 3. Calculate the loss function by comparing the model's inference results with the true labels;

[0091] Step 4. Perform gradient backpropagation based on the loss function to update the model weights. If training is not finished, return to step 2; if training is finished, proceed to step 5.

[0092] Step 5. Apply the trained multimodal fusion model to the test set and obtain the classification metric.

[0093] refer to Figure 6 The workflow of the non-invasive coronary heart disease detection system in this embodiment for coronary heart disease screening is as follows:

[0094] E1. Pre-filter the 12-lead ECG;

[0095] E2.1 feeds the filtered ECG into a one-dimensional convolutional network and a Transformer encoder for feature extraction;

[0096] E2.2 performs a short-time Fourier transform on the filtered ECG to generate a two-dimensional image, which is then fed into the Vit network for feature extraction.

[0097] E2.3 performs Kors regression on the filtered ECG to generate VCG signals, and then extracts features through a one-dimensional convolutional network and a Transformer encoder.

[0098] E2.4 extracts information such as age, weight, and gender from electronic medical records and feeds it into a text encoder to obtain text features;

[0099] E3. The four features are fused using a self-attention mechanism;

[0100] E4. Use the fused features for initial screening of coronary heart disease.

[0101] This embodiment utilizes data modeling from electrocardiogram (ECG) vector maps. The ECG is obtained by projecting the ECG vector map twice. Adding information from the ECG vector map can greatly compensate for the information lost in the ECG. The multimodal feature extraction and fusion model structure can integrate information from ECG, ECG vector maps, spectrograms, and electronic medical records. The model structure of the ECG signal feature extraction unit can use branching structures and CNNs to extract low-level features, and transformers to extract high-level features. This establishes the temporal correlation of ECG signals and uses high-level features for data augmentation, making the prediction results more accurate.

[0102] Experimental example:

[0103] 1. Dataset

[0104] 1.1 Public Dataset PTB Diagnostic ECG Database

[0105] The PTB Diagnostic ECG Database is a large and rich dataset containing electrocardiogram (ECG) records from 290 patients, including 148 diagnosed with myocardial infarction and 52 healthy individuals, aged 15 to 88 years. Each record was sampled at 1000 Hz for 10 seconds. Each record includes 12-lead ECG signals and 3-lead VCG signals, with the ECG vectorgraphs providing richer ECG features for model training and testing. The database has been manually reviewed and annotated, with detailed clinical annotations for each record, including patient history, physical examination results, and ECG findings. Due to its large volume of data from myocardial infarction patients, this database is widely used in research on myocardial infarction detection and ECG signal processing. Coronary artery disease (CAD) and myocardial infarction are two closely related cardiovascular diseases. CAD is a heart disease caused by insufficient blood supply to or blockage of the coronary arteries, with common symptoms including chest pain, shortness of breath, and palpitations. Myocardial infarction is caused by myocardial ischemia and necrosis due to coronary artery obstruction, and usually manifests as severe chest pain, palpitations, and shortness of breath. Therefore, myocardial infarction can be considered a serious complication of coronary heart disease. We therefore used the PTB Diagnostic ECG Database dataset to validate the model's performance.

[0106] 1.2 Clinical Data Collection Database: PKUSZ Diagnostic CAD Database

[0107] Currently, there are relatively few publicly available coronary artery disease (CAD) datasets in China, and both the quantity and quality of the data need improvement. This experimental example uses the high-quality clinically acquired ECG database, PKUSZ Diagnostic CAD Database. This dataset consists of 1153 ECG records, including 457 CAD patients, 384 patients with arrhythmias, and 312 healthy controls. Each ECG record contains 12 leads (I, II, III, aVR, aVR, aVR, aVR, aVR, aVR, V1, V2, V3, V4, V5, and V6) with a sampling rate of 500 Hz. This dataset was correctly diagnosed and annotated by doctors at a certain hospital, containing rich clinical annotations and diagnostic information, making the data quality more reliable. Establishing this dataset can provide more accurate and abundant data resources for CAD research and diagnosis. The dataset covers a large number of ECG records from CAD patients, which helps in-depth research into the pathophysiological mechanisms and diagnostic methods of CAD, providing better support for clinical diagnosis and treatment.

[0108] 2. Implementation Details

[0109] All experiments in this paper are implemented using PyTorch. In the model parameter settings, the proposed model was trained for 70 epochs with a batch size of 16. This experimental example uses the Adam optimizer with a learning rate of 0.0001. The learning rate employs a fixed-step decay method, decreasing to 0.9 every 10 epochs. ReLU is used as the default activation function, and softmax is used as the loss function. To mitigate overfitting, an early stopping criterion is adopted; training stops when the validation F1 score does not improve for more than 20 iterations. The best-performing model is used for performance evaluation.

[0110] 3. Evaluation Indicators

[0111] In this study, model performance on the dataset was evaluated using four metrics: sensitivity (Sen), specificity (Spe), accuracy (Acc), and F1-score. Sen, also known as recall, represents the percentage of correctly predicted positive samples out of all positive samples. Conversely, Spe represents the percentage of correctly predicted negative samples out of all negative samples, and accuracy (Acc) is the percentage of all correctly identified samples out of the total sample count. The F1-score was primarily used to evaluate recognition performance and is the most important evaluation metric for both datasets.

[0112] 4. Experimental Results

[0113] 4.1 The comparative experimental results on the public dataset are shown in Table 1 below. Table 1 compares the performance of the model in this embodiment with the existing technology model based on the PTB data test set:

[0114] Table 1

[0115]

[0116] As can be seen, MF-CADNet performs excellently, outperforming other models in all evaluation metrics. The non-invasive coronary artery disease detection system in this experiment achieved an F1 score of 98.41%, a 0.11% improvement over EvoMBN's F1 score and a 0.28% improvement in accuracy. Sensitivity and specificity both exceeded 95%. Furthermore, most existing technologies use single-lead or 12-lead data, utilizing only a portion of the database information. MF-CADNet in this experiment incorporates lead information from the VCG, providing a more comprehensive view of cardiac electrical activity and effectively improving accuracy. Most existing models use CNNs as the network backbone and RNNs to extract temporal information. Considering the limitations of CNN receptive fields and the poor parallelism of RNNs, MF-CADNet in this experiment employs a combined CNN and transformer encoder structure, enabling the extraction of long-range feature information.

[0117] 4.2 Experimental results on clinical datasets are shown in Tables 2 and 3 below. Table 2 evaluates the model's performance in predicting F1 scores for each disease category based on a test set of coronary heart disease data from a certain hospital. Table 3 compares the performance of this experimental case model with other existing technology models based on a test set of coronary heart disease data from a certain hospital.

[0118] Table 2

[0119]

[0120] Table 3

[0121]

[0122] For this three-class classification problem, the MF-CADNet in this experimental example outperforms existing models in all metrics. Regarding the number of leads used, most existing models employ 12-lead ECG. In contrast, this model uses 15-lead ECG (12-lead ECG and 3-lead VCG), and incorporates spectrograms and personalized patient information, significantly enhancing the richness of ECG data. Table 2 lists the F1 scores predicted for each class by each model. Here, cad represents coronary artery disease, arrhythmia represents arrhythmia, and hc represents the healthy control group. The F1-score of MF-CADNet in this experimental example is higher than 91, indicating that the system can accurately classify patients with coronary artery disease, patients with arrhythmia, and normal individuals. However, other models can only distinguish between diseased and healthy individuals, but cannot further differentiate between coronary artery disease and other arrhythmias. In terms of sensitivity, specificity, and accuracy, the system in this experimental example performs exceptionally well, all reaching state-of-the-art (SOTA) levels. In terms of model structure, compared with the CNN-LSTM network structure in RC-NNs, the experimental example of this invention uses Vit to combine the advantages of CNN and Transformer, deepens the network, and improves the feature extraction capability by utilizing the self-attention mechanism.

[0123] 4.3 Ablation Experiment

[0124] To analyze the relative contributions of different modules in the model, we evaluated variations of the experimental model with different settings. Refer to Tables 4 and 5 below. Table 4 evaluates the performance of the ECG vector map feature extraction unit based on a coronary heart disease data set from a certain hospital, and Table 5 evaluates the performance of the inter-module attention mechanism based on the same data set.

[0125] Table 4

[0126]

[0127] As shown in Table 4, the classification performance of the model in the experimental examples was effectively improved by adding VCG information. The performance of the model without VCG information and the model with VCG information were evaluated in the experimental examples. We found that the model with VCG information had better performance, with improvements in all evaluation metrics, including an increase of approximately 6.5% in F1. The experimental results demonstrate that the ECG vector map feature extraction unit can effectively supplement information and improve classification accuracy.

[0128] Table 5

[0129]

[0130] To evaluate the effectiveness of the attention-based fusion module in this experimental example, additional experiments were conducted using a model without the attention fusion module. In this experiment, the attention fusion module was removed before the classifier, while other network configurations remained unchanged during training. The experimental results are shown in Table 5. The attention-based fusion module increased the average sensitivity by 2.1% and the F1 score by approximately 2.3%. This indicates that the attention-based fusion module helps to uncover discriminative features and effectively improves the model's classification performance.

[0131] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0132] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0133] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0134] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0135] The above description provides a further detailed explanation of the present invention in conjunction with specific / preferred embodiments, and it should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various substitutions or modifications can be made to these described embodiments without departing from the concept of the present invention, and all such substitutions or modifications should be considered within the scope of protection of the present invention. In the description of this specification, the reference to terms such as "an embodiment," "some embodiments," "preferred embodiment," "example," "specific example," or "some examples," etc., indicates that the specific features, structures, materials, or characteristics described in connection with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any suitable manner in one or more embodiments or examples. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification and the features of different embodiments or examples. Although the embodiments of the present invention and their advantages have been described in detail, it should be understood that various changes, substitutions, and modifications can be made herein without departing from the scope of protection of the patent application.

Claims

1. A non-invasive coronary heart disease detection system, characterized in that, include: A preprocessing module is used to filter out noise from the electrocardiogram signal and extract relevant information about coronary heart disease from the electronic medical records associated with the electrocardiogram signal; The multimodal feature extraction module includes an electrocardiogram signal feature extraction unit, a time-frequency image feature extraction unit, an electrocardiogram vector map feature extraction unit, and an electronic medical record text feature extraction unit; The system is used to receive the preprocessed electrocardiogram (ECG) signal and related information about coronary heart disease, and extract ECG signal features, time-frequency image features, ECG vector map features, and electronic medical record text features. The ECG signal feature extraction unit is used to obtain different features from the ECG signal by passing it through at least two feature extractors of different scales, then concatenate and fuse the different features, and finally obtain the ECG signal features through a channel attention mechanism and a residual structure, and finally through an adaptive attention mechanism. The time-frequency image feature extraction unit is used to perform a short-time Fourier transform on the ECG signal to generate a two-dimensional image and then perform feature extraction through a Vit network to obtain the time-frequency image features. The ECG vector map feature extraction unit is used to generate an ECG vector map signal from the ECG signal through Kors regression, and then perform feature extraction through a one-dimensional convolutional network and a Transformer encoder to obtain the ECG vector map features. The multimodal feature fusion module is used to concatenate and fuse the electrocardiogram signal features, the time-frequency image features, the electrocardiogram vector map features, and the electronic medical record text features through a cross-module attention mechanism to obtain fused features; The ECG classification module is used to predict the labels of the fused features to obtain the detection results; The system integrates information from electrocardiograms, spectrograms, vectorcardiograms, and electronic medical records, extracts features from different signal segments using convolutional kernels with different receptive fields, and extracts richer features using inter-lead attention mechanisms and inter-module attention mechanisms.

2. The non-invasive coronary heart disease detection system as described in claim 1, characterized in that, The multimodal feature fusion module fuses the electrocardiogram signal features, the time-frequency image features, the electrocardiogram vector map features, and the electronic medical record text features using a self-attention mechanism.

3. The non-invasive coronary heart disease detection system as described in claim 1, characterized in that, The ECG signal feature extraction unit structure is one of the following network structures: convolutional neural network, a combination of CNN and LSTM, a combination of CNN and Transformer, or Transformer.

4. The non-invasive coronary heart disease detection system as described in claim 1, characterized in that, The test results indicated one of the following: coronary heart disease, arrhythmia, or good health.

5. The non-invasive coronary heart disease detection system as described in claim 1, characterized in that, The preprocessing module uses at least one of the following methods to filter out ECG signal noise: bandpass filter, least mean square adaptive filter, wavelet transform-based filter, statistical model-based filter, neural network-based filter, maximum-minimum normalization, and sigmoid normalization.

Images