Publication

Abstract : Cardiovascular diseases (CVDs) continue to pose a major global health challenge, playing a significant role in premature mortality. An electrocardiogram (ECG) stands out as one of the most accessible and non-invasive diagnostic tools for detecting cardiac abnormalities. Manual ECG interpretation takes a lot of time and is prone to mistakes, especially in busy clinical environments. This emphasizes the importance of precise and automated diagnostic solutions. This study introduces a deep learning framework designed for ECG image classification, utilizing pre-trained convolutional neural networks (CNNs) alongside a unique hybrid model. Three established CNN architectures—EfficientNetB0, InceptionV3, and ResNet50—were assessed using a two-stage approach: feature extraction with frozen backbones, followed by fine-tuning. A hybrid model was developed that combines ResNet50 with Bidirectional Long Short-Term Memory (BiLSTM) layers and an attention mechanism to effectively capture spatial, temporal, and contextually significant patterns in ECG waveforms. A stratified 70-15-15 data split with robust preprocessing ensured data quality, while training techniques like dropout, class weighting, and dynamic learning rate improved model performance. The hybrid model achieved 93.57% validation accuracy and 82% class-wise accuracy, outperforming baselines. These results underscore the potential of combining spatial, sequential, and attention-based learning for scalable cardiovascular diagnosis