SACNN: a spatial attentive 2D CNN for improved accuracy and interpretability in ECG image classification

Deep Learning models have shown strong performance in ECG analysis. However, many existing models operate as black-box systems or depend on post-hoc explainable techniques, limiting transparency. Moreover, attention-based mechanisms have focused on 1D ECG signals with limited exploration of ECG images. In this research, we propose a Spatial Attention-based 2D Convolutional Neural Network (SACNN) with an intrinsic attention mechanism that enhances model transparency by highlighting regions emphasized during feature extraction. The incorporation of Spatial attention into the 2D-CNN within the learning process aided in visualizing image regions emphasized by the model without requiring post-hoc explainability methods. The experimental analysis showed classification performance with an accuracy of up to 0.98 on a single split and consistent performance with a mean accuracy of 0.93 under stratified k-fold cross-validation. Statistical hypothesis testing confirmed that the performance improvements of SACNN over the evaluated baseline CNN models were statistically significant (p<0.05) and associated with large effect sizes. Furthermore, the model maintains its attention on waveform-dominant areas over background regions and exhibits robustness under noise perturbations. The quantitative analysis of attention distribution in the heatmaps, supported by expert assessment, indicates that the learned attention behavior is spatially reasonable. This work represents a foundational step toward transparent deep learning frameworks for ECG image analysis and highlights the potential of intrinsic attention heatmaps to improve transparency and facilitate visual assessment of model attention.