Lightweight Data-driven ECG Classification Approach with Explainable CAM Output

This study proposes a lightweight, data-driven approach for classifying electrocardiogram (ECG) signals with enhanced explainability using a non-trainable Class Activation Map (CAM) as a second output. The model, built on a minimalistic and streamlined convolutional neural network (CNN) architecture, provides binary classification (normal versus disease-related signals) while concurrently generating an interpretable overlapping map of signal regions responsible for model decisions. The CAM output serves as a lightweight enhancement, requiring no additional trainable parameters and introducing only a 14.99% increase in computational complexity compared to the proposed classification model. The model remains highly efficient with just 22,273 parameters and 92,064 floating-point operations (FLOPs) while producing classification and explainability outputs. Trained and evaluated on the PTB-XL dataset, it achieves an accuracy of 85.4% and an area under the ROC curve (AUC) of 0.93. The data preprocessing pipeline includes detrending, standardization, and augmentation techniques. By producing both classification results and explanatory heatmaps in a single computational pass, the model reduces complexity while enhancing clinical interpretability. This approach demonstrates the feasibility of deploying explainable AI models for real-time cardiac anomaly detection, advancing trustworthy and efficient diagnostic systems in healthcare environments. The study aims to contribute to the advancement of explainable artificial intelligence and its integration into clinical decision-making frameworks.