Integrating multimodal information in machine learning for classifying acute myocardial infarction

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Prompt identification of myocardial infarction (MI) is critical in acute coronary syndrome management.
• Electrocardiograms (ECGs) are initial screening tools but have limited diagnostic accuracy.
• Existing machine learning (ML) models often fail to account for MI-confounding conditions and underutilize clinical data.

Purpose of the Study:

• To develop and evaluate ML models for MI detection, considering clinical implementation scenarios.
• To investigate the impact of ECG duration and multimodal information on model performance.
• To propose a novel multimodal deep learning architecture integrating ECG and patient demographics.

Main Methods:

• A large-scale ECG dataset was dichotomized into MI and non-MI classes, including MI-confounding conditions.
• Two experiments assessed the influence of ECG duration and multimodal data on model training.
• A multimodal deep learning architecture was developed to learn joint features from ECG and patient demographics.

Main Results:

• The multimodal model demonstrated superior performance over ECG-only models, achieving 92.1% AUC and 87.4% accuracy.
• Performance was comparable to existing studies despite increased task difficulty from the new class definition.
• Model explainability confirmed the contribution of patient demographics and clinical concordance.

Conclusions:

• Multimodal deep learning models integrating ECG and patient demographics show significant promise for accurate MI detection.
• These findings support the development of ML solutions for prompt MI diagnosis, advancing clinical application.
• The study highlights the importance of considering confounding conditions and multimodal data in ML model design for cardiology.