Research And System Implementation Of Myocardial Infarction Diagnosis Based On Multimodal Data

The incidence of acute myocardial infarction is a dangerous condition that can lead to numerous complications,and it tends to occur more frequently in younger individuals.However,delayed diagnosis of myocardial infarction can result in a significant delay in patients receiving timely medical attention.Issuing early warning signals through regular electrocardiographic examinations would be of great assistance in the early detection of myocardial infarction,facilitating prompt treatment and ultimately improving patient survival rates.Previous research on auxiliary diagnosis of acute myocardial infarction using ECGassisted diagnosis technology has typically focused on analyzing a single type of data and establishing a direct mapping from data to results.However,relying solely on ECG signals may not provide sufficient information for accurately predicting these situations or events.Therefore,the utilization of multimodal information fusion technology becomes crucial,as it can integrate different types of information to enhance the performance of the diagnostic model.The main objective of this thesis is to address the auxiliary diagnosis algorithm of myocardial infarction based on electrocardiogram(ECG)and its practical application.This research employs signal processing technology to effectively filter out common noise in the dataset,addresses the problem of unbalanced sample distribution that often leads to misclassification,and integrates multiple modal ECG data for comprehensive cardiology analysis.By achieving these objectives,this thesis aims to contribute to the early detection of myocardial infarction,facilitate prompt treatment,and ultimately improve patient survival rates.(1)To support further research,the collected ECG dataset was transformed from picture format to time series signal data using Paper ECG.To address common issues in ECG signals like baseline drift,muscle artifacts,and power line interference,a combination of low-pass filter(FIR)and high-pass filter(IIR)was utilized to effectively eliminate the noise present in the signal.This filtering procedure enhances the quality of the ECG data and enables more precise analysis in subsequent research.(2)In this thesis,the primary objective is to address the challenge of class imbalance arising from limited samples.To tackle this issue,an optimization-based meta-learning method is proposed,which utilizes a bidirectional long-term short-term memory neural network.The aim is to learn reweighting functions adaptively using a small unbiased dataset.The ultimate goal is to achieve a balanced correlation between electrocardiographic signals from 12 leads and different types of myocardial infarction.The experimental results demonstrate significant improvements in performance compared to the baseline model when dealing with unbalanced ECG data.Furthermore,the proposed method exhibits enhanced representation capabilities for myocardial infarction types that do not show obvious changes in ECG bands.(3)This study addresses the issue of single-modal ECG signals potentially lacking accurate information by utilizing the knowledge distillation mechanism.In the pre-training task,a teacher model is trained using a combination of a 1D neural network for time series data and a 2D neural network for spectrogram data.Subsequently,a lightweight student model is employed to fuse the features extracted from the two networks using a gate fusion mechanism.The results of the ablation experiment demonstrate the positive impact of both the gate fusion mechanism and knowledge distillation on performance improvement.The study emphasizes the advantages of utilizing multimodal data compared to single-modal data.Finally,the prediction results of the multimodal model are interpreted visually through the SHAP method,which provides insights into the decision-making process of the model.(4)An auxiliary diagnosis platform for myocardial infarction is designed and implemented in this thesis.Building upon the multimodal myocardial infarction auxiliary diagnosis model,an automated ECG detection application is developed.The system enables the import of ECG signals,which are then processed by the signal processing module and feature extraction module.The processed signals are displayed in an interactive interface,allowing for ECG visualization.Additionally,the system performs feature extraction and provides auxiliary diagnosis for myocardial infarction.Furthermore,a case table functionality is included to facilitate the management of patient cases.Thorough experiments were conducted in this thesis using electrocardiogram data collected from two tertiary hospitals in Xinjiang.The experimental results unequivocally establish the superior performance of the multimodal auxiliary diagnosis model for myocardial infarction.The accuracy of myocardial infarction recognition reached 86.27%,surpassing the single-modal model by 1.27%.Additionally,the myocardial infarction auxiliary diagnosis platform developed in this thesis adequately meets the initial clinical requirements.