基于语义引导与多头交叉融合的心梗检测模型OA

Timely detection of Myocardial Infarction(MI)is crucial for reducing mortality rates and informing clinical decision-making.Existing deep learning-based MI detection methods have begun to adopt multimodal approaches to enhance performance.However,some current methods that concatenate ECG waveforms and diagnostic texts may amplify semantic noise and fail to adaptively balance modal contributions.Meanwhile,certain methods applying global alignment paradigms in the cardiovascular domain to obtain general representations struggle to capture the specific aligned representations between MI ECG waveforms and textual semantics,making it difficult to effectively distinguish between ST-segment elevation myocardial infarction(STEMI)and its mimics—such as angina,pericarditis,and hyperkalemia—particularly in the presence of textual noise,while maintaining interpretability.To address these issues,we propose a myocardial infarction detection model named SCAFDM,based on semantic guidance and multi-head cross-attention fusion.The model designs a multi-head cross-attention mechanism to align ECG waveforms and diagnostic texts,dynamically assigning weights to text tokens to suppress semantic noise and achieve adaptive inter-modal balance.Additionally,an end-to-end joint loss training method is adopted to resolve the feature alignment problem in multimodal MI detection.Experiments on the PTB-XL dataset demonstrate that SCAFDM achieves excellent performance in terms of area under the ROC curve and F1 score.