Optimal ECG-lead selection increases generalizability of deep learning on ECG abnormality classification

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning (DL) shows promise for 12-lead electrocardiogram (ECG) abnormality detection.
• Redundancy in 12-lead ECG data can lead to DL model overfitting and poor generalization.
• Optimizing lead subsets may enhance DL model performance in ECG interpretation.

Purpose of the Study:

• To develop and evaluate a DL model with an integrated ECG-lead subset selection stage.
• To identify an optimal subset of ECG leads that minimizes redundancy and improves model generalizability.
• To automatically interpret multiple common ECG abnormality types using an optimized lead subset.

Main Methods:

• A DL model incorporating feature extraction, ECG-lead subset selection, and decision-making stages was developed.
• The model was trained and validated on CPSC 2018 (6877 recordings) and tested on PhysioNet/CinC 2020 (3998 recordings).
• An ECG-lead subset selection module identified an optimal 4-lead subset: leads II, aVR, V1, and V4.

Main Results:

• The DL model utilizing the optimal 4-lead ECG subset significantly outperformed the 12-lead model on validation and external test datasets.
• The identified 4-lead subset (II, aVR, V1, V4) effectively reduced data redundancy.
• The proposed model demonstrated improved generalizability in interpreting ECG abnormalities.

Conclusions:

• An optimal 4-lead ECG subset can enhance the generalizability of DL models for ECG abnormality interpretation.
• This finding offers insights into essential ECG leads for automated cardiac abnormality detection systems.
• Reducing data redundancy through lead selection is crucial for robust DL-based ECG analysis.