EXGnet: a single-lead explainable-AI guided multiresolution network with train-only quantitative features for trustworthy ECG arrhythmia classification

Explainability improves clinical trustworthiness: EXGnet uses XAI (explainable AI) supervision during training to direct the model's attention to clinically relevant ECG regions, mimicking how cardiologists analyze heart rhythms. This approach helps make the AI's decision-making process transparent and aligned with medical expertise, addressing a key barrier to clinical adoption.

Automatic feature generation eliminates manual annotation burden: The model generates its own ground truth for explainability training using a heart rate variability-based approach, removing the need for manual annotation by cardiologists. This scalability advantage could accelerate deployment in resource-limited healthcare settings.

Smart feature engineering balances accuracy and deployment: EXGnet incorporates quantitative ECG features during training to improve accuracy (achieving 98.762% on Chapman dataset), but excludes them during inference to keep the model lightweight for edge devices. This "train-only" approach demonstrates how to maintain high performance without sacrificing practical deployability.

Single-lead ECG accessibility expands potential reach: By working with single-lead signals rather than full 12-lead ECGs, EXGnet could enable arrhythmia screening on simpler, more accessible devices like smartwatches or portable monitors, potentially democratizing cardiac monitoring.

Rigorous benchmarking demonstrates real-world readiness: Validation on two independent benchmark datasets (Chapman and Ningbo) with high F1-scores (97.910% and 95.527%) provides evidence of generalization beyond a single dataset, supporting confidence in clinical applicability.