Bridging clinical knowledge and AI: an interpretable transformer framework for ECG diagnosis

Area of Science:

• Artificial Intelligence in Medicine
• Cardiology
• Biomedical Signal Processing

Background:

• Cardiovascular diseases represent a significant global health challenge.
• Accurate, efficient, and interpretable automated electrocardiogram (ECG) interpretation is crucial.
• Existing systems often lack physician-level interpretability.

Purpose of the Study:

• To develop MDOT (Momentum Distillation Oscillographic Transformer), an AI model for automated ECG interpretation.
• To enhance diagnostic accuracy and clinical interpretability of ECG analysis.
• To provide a robust solution for clinical decision support and edge deployment.

Main Methods:

• A lightweight student model trained using knowledge distillation from a physician-informed teacher model.
• Integration of key electrophysiological indicators (heart rate, QRS duration, ST segment, QTc interval) as auxiliary features.
• A novel Oscillographic (OSC) module converting 1D ECG to 2D representations for Transformer analysis.
• Utilizing attention mechanisms to generate interpretability heatmaps.

Main Results:

• MDOT achieved state-of-the-art accuracy on two benchmark datasets: 99.53% on MIT-BIH (8 classes) and 99.03% on Chapman (12 classes).
• The model demonstrated high performance on strict inter-patient data splits.
• Attention-based heatmaps highlighted diagnostically relevant ECG segments, improving interpretability.

Conclusions:

• MDOT offers a highly accurate and interpretable automated ECG interpretation system.
• The integration of clinical reasoning and novel signal processing enhances diagnostic capabilities.
• MDOT presents a promising tool for clinical decision support and real-world applications.