Interpretable arrhythmia detection in ECG scans using deep learning ensembles: a genetic programming approach.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Informatics

Background:

• Cardiovascular diseases are a leading cause of death globally.
• Accurate arrhythmia detection and atrial fibrillation (AF) recurrence prediction are critical for patient outcomes.
• Current diagnostic methods can be limited in speed and accuracy.

Purpose of the Study:

• To develop and validate deep learning ensembles for arrhythmia detection and AF recurrence prediction using electrocardiogram (ECG) data.
• To enhance model interpretability through explainable artificial intelligence (XAI).
• To compare the performance of ensemble models against individual and voting models.

Main Methods:

• Utilized deep learning ensemble models trained on ECG data from two large patient cohorts (Dataset G and Dataset L).
• Employed explainable artificial intelligence (XAI) techniques to provide insights into model decision-making.
• Validated model performance using metrics including Area Under the Receiver Operating Characteristic Curve (ROC-AUC) and Precision-Recall AUC.

Main Results:

• Deep learning ensembles demonstrated superior performance in arrhythmia detection compared to individual models.
• Achieved high ROC-AUC values: 0.980 for Dataset G and 0.799 for Dataset L.
• Ensemble models showed improved Precision-Recall AUC for AF recurrence prediction (0.765) versus individual models (0.737).

Conclusions:

• Deep learning ensembles offer a powerful tool for arrhythmia detection and AF recurrence prediction from ECGs.
• XAI integration improves the clinical applicability and trustworthiness of AI-driven cardiovascular diagnostics.
• The findings support the potential of AI to enhance the management of cardiovascular diseases.