Modified ResNet-50 and Custom CNN Ensembles Achieve Near Perfect STEMI Detection on European ST-T Database

Background ST-segment elevation myocardial infarction (STEMI) affects over 750,000 Americans annually, with 30-day mortality rate of 2.5–10%. While traditional risk factors include hypertension, hyperlipidemia, smoking, and diabetes, though SMuRF-less cases suggest genetic or non-atherosclerotic causes. Rapid diagnosis via electrocardiogram (ECG) ST- elevation and troponin levels is vital, requiring primary percutaneous coronary intervention (PCI) within 120 minutes, supported by dual antiplatelet and anticoagulant therapy. Recent advancements (2023–2025) in AI-enhanced ECG interpretation, PCSK9 inhibitors, and revascularization highlight the need for rapid detection. Methods Our study introduces a hybrid model for automated ECG classification of ST-segment abnormalities, including STEMI and ischemia, using the European ST-T Database. The base model, Modified ResNet-50 with a sparse autoencoder, is optimized via an ensemble approach and compared with MobileNetV2, LeNet5, and Custom CNN, all employing ensemble classification. The dataset comprised 10,800 training and 2,700 testing samples, despite class imbalance. Results Modified ResNet-50 with Ensemble achieved 99.78% accuracy, 95.12%-99.92% precision, 97.50%-99.85% recall, 98.67% AUC, and 96.19% MCC. Custom CNN with Ensemble excelled with 99.78% accuracy, 98.76%-99.81% precision, 94.12%-99.96% recall, 97.04% AUC, and 96.30% MCC. MobileNetV2 and LeNet5 with Ensemble scored lower across metrics. Conclusion The proposed models enable rapid STEMI detection, enhancing PCI timeliness, telehealth monitoring, and AI-assisted ECG training. Custom CNN with Ensemble offers superior precision and MCC, ideal for edge devices, while Modified ResNet-50 ensures high-sensitivity diagnostics, forming a robust framework for improved cardiac care.