A Comparative Benchmark of Deep Learning Models and Deployment of a Web Application for Automated Early Heart Attack Risk Prediction

Abstract: Heart attacks remain one of the leading causes of death worldwide, highlighting the importance of early and accurate prediction. This study focuses on developing and comparing two deep learning models—CNN + MobileNet and EfficientNetV2B3—for automatic classification of ECG images and deploying the best-performing model as a real-time web application. An ECG image dataset consisting of 1,377 samples across four classes (Normal, Myocardial Infarction, Abnormal Heartbeat, and History of MI) was obtained from Kaggle. Both models were trained using TensorFlow–Keras with data augmentation and hyperparameter tuning on Google Colab (Tesla T4 GPU). Their performance was evaluated using accuracy, precision, recall, and F1-score metrics. Results showed that the CNN + MobileNet model outperformed EfficientNetV2B3, achieving 89% accuracy, 0.89 precision, 0.88 recall, and 0.89 F1- score, compared to 79% accuracy for EfficientNetV2B3. Additionally, CNN + MobileNet demonstrated smoother convergence, faster inference time (~150 ms per image), and a lightweight model size (~30 MB). Thus, CNN + MobileNet proved to be more effective for ECG classification and real-time prediction, and its deployment through a Gradio-based web interface enables accessible and rapid heart attack detection, especially in remote healthcare settings.

Keywords: Heart attack prediction, ECG classification, Deep learning, CNN + MobileNet, EfficientNetV2B3, Web deployment.