A Comparative Study of Clinical Rule-Based and Deep Learning-Based Diagnosis for Myocardial Infarction Detection Using Electrocardiograms

Myocardial infarction (MI) diagnosis relies on established clinical criteria, primarily involving the identification of electrocardiographic (ECG) abnormalities, such as ST-segment elevation in anatomically contiguous leads. Although rule-based algorithms grounded in these guidelines remain prevalent in clinical practice, recent advances in deep learning (DL) have demonstrated promising performance. However, there is a lack of studies comparing these two methods. To this aim, we developed a multitask DL model and evaluated its performance in comparison with a rule-based algorithm for MI detection, anatomical localization (anterior, lateral, inferior, and septal), and stadium classification (acute, chronic, and normal). The model was trained using 12-lead median beats from the PTB-XL+ dataset. The DL model achieved a sensitivity (Se) of 0.89 and a specificity (Sp) of 0.96 for MI detection, outperforming the rule-based algorithm, which achieved a Se of 0.69 and a Sp of 0.94. For MI localization, the DL model achieved an average F1 score across regions of 0.72, while the rule-based algorithm 0.55. In MI stadium classification, the DL model attained an average F1 score of 0.68, compared to 0.58 for the rule-based method. Overall, the DL model outperformed the rule-based algorithm across all tasks.