Evaluation of customized adaptive cardiovascular activation function with deep neural networks for myocardial infarction classification using ECG image

Myocardial infarction (MI), in its early or marginal stages, often presents with subtle and complex electrocardiogram (ECG) patterns that are frequently misclassified or overlooked by conventional diagnostic methods. Traditional ECG analysis techniques rely heavily on handcrafted features, rule-based thresholds, pretrained models, and domain-specific preprocessing, rendering them inadequate for precise, timely myocardial detection in real-time clinical workflows. To address these challenges, this research proposes a robust deep learning-based model with a customized activation function for the automatic classification of MI ECG signals, with a primary focus on accurate identification of myocardial cases, along with Normal and other Abnormal cases. The model combines a ResNet-50 architecture fused with a customised activation function for improved feature extraction. This model is integrated with a Bayesian-optimised Random Forest classifier to ensure accurate and generalised classification. The proposed model achieves an accuracy of 99.9%, precision of 99%, recall of 99%, and specificity of 98% by surpassing the performance of adaptive cardiovascular activation function (ACAF)-based Residual Network (ResNet) and other baseline methods. Comparative interpretation with existing models confirms its higher performance and reliability. These results indicate that the proposed approach can serve as a potent, deployable diagnostic system for automated detection of early myocardial infarction.