Fusion CNN and ViT with self-Attention for Multi-Label ECG Classification

Abstract

Cardiovascular diseases remain the leading cause of mortality worldwide, underscoring the critical importance of accurate and timely diagnosis. Electrocardiography (ECG) is a widely used, non-invasive technique for detecting cardiac abnormalities, yet its manual interpretation remains challenging—particularly in multi-label scenarios where several co-occurring conditions may be present in a single recording. This thesis addresses these challenges by investigating a range of machine learning (SVM, Random Forest, XGBoost) and deep learning architectures (CNN, LSTM, Transformer), with a particular focus on multi-label ECG classification using the PTB-XL dataset. We propose a hybrid deep learning model combining Convolutional Neural Networks (CNN), Transformer layers, and a Self-Attention mechanism. This architecture was specifically designed to improve the classification of both common and underrepresented cardiac conditions while maintaining computational efficiency. Experimental results demonstrate that the proposed hybrid model outperforms both standalone deep learning and classical machine learning models. It achieved a Binary Accuracy of 93.32%, a Micro F1-score of 0.7663, and a recall of 0.7238, while reducing false negatives in classes such as Myocardial Infarction (MI) and Hypertrophy (HYP). In addition, the model integrates Focal Loss to mitigate class imbalance and employs explainability techniques (XAI) to enhance interpretability in clinical applications. These findings confirm the potential of hybrid architectures in delivering robust, scalable, and interpretable solutions for real-world ECG classification tasks.