針對心電圖資料不平衡之分類模型設計

Abstract

本研究旨在探討運用深度學習技術於心電圖（ECG）訊號分類的應用潛力，以協助提升心律異常的辨識能力與早期診斷準確性。研究中提出一種基於一維殘差網路（1D ResNet-18）之模型架構，並整合卷積區塊注意力模組（CBAM）與輔助分類器（Auxiliary Classifier），以強化模型對 ECG 特徵的表達與判別能力。此架構源自電腦視覺任務，經調整後應用於一維生理訊號的分類工作，展現良好的適應性。資料處理方面採用 ADASYN 技術處理類別不平衡問題，並輔以資料增強策略以提升模型穩定性與泛化能力。模型於 MIT-BIH 公開資料集中進行驗證，結果顯示其分類表現優於傳統方法，特別是在多類別訊號辨識上具備一定的穩定性與準確性。綜合研究結果，顯示本模型結合注意力機制、輔助分類設計與資料處理策略後，能有效強化 ECG 訊號分類模型之應用能力，未來有望作為智慧健康照護輔助診斷系統的技術參考。

This study investigates the use of deep learning in classifying electrocardiogram (ECG) signals to support arrhythmia detection and early diagnosis. A modified one-dimensional ResNet-18 model is introduced, integrating a Convolutional Block Attention Module (CBAM) and an auxiliary classifier to enhance feature representation and classification. These mechanisms, originally designed for computer vision tasks, are adapted to one-dimensional biomedical signals with promising results. To address class imbalance, the Adaptive Synthetic Sampling Approach (ADASYN) and data augmentation are applied, improving training stability and generalization. Using the MIT-BIH public dataset, the model demonstrates more consistent and accurate performance than conventional methods, particularly for underrepresented classes.The combination of attention mechanisms, auxiliary classification, and data balancing strategies establishes a robust ECG classification framework with potential applications in intelligent healthcare and clinical decision support.