Signal-Guided Multitask Learning for Myocardial Infarction Classification Using Images of Electrocardiogram

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Electrocardiogram (ECG) interpretation for myocardial infarction (MI) diagnosis is challenging.
• Accurate and swift MI diagnosis is critical in clinical practice.
• Developing advanced diagnostic tools for MI is essential.

Purpose of the Study:

• To develop a deep learning (DL) algorithm for differentiating MI from non-MI ECGs.
• To utilize a signal-guided multitask learning approach for enhanced ECG interpretation.
• To evaluate the DL algorithm's performance in a clinical setting.

Main Methods:

• A DL model was trained on 11,227 ECG images using a signal-guided multitask learning framework.
• The algorithm was built upon a prior single-task DL model.
• The DL model's impact on diagnostic accuracy was assessed by having 51 physicians interpret ECGs with and without AI assistance.

Main Results:

• The DL algorithm achieved high performance metrics: 90.56% accuracy, 83.82% sensitivity, 93.02% specificity, 81.44% precision, and an 82.61% F1 score.
• Physician accuracy in ECG interpretation improved from a median of 62% to 68% with DL assistance.
• Significant accuracy gains were observed in both internal medicine trainees and specialists, and the algorithm improved interpretation for both STEMI and NSTEMI cases.

Conclusions:

• The signal-guided multitask DL algorithm significantly outperforms previous single-task models in discriminating MI ECGs.
• The DL algorithm serves as a valuable decision support tool for physicians diagnosing MI.
• The diagnostic improvements provided by the DL algorithm are consistent across STEMI and NSTEMI subgroups.