From Prototypes to Sparse ECG Explanations: SHAP-Driven Counterfactuals for Multivariate Time-Series Multi-class Classification

Sparse explanations improve clinical usability: The framework modifies only 78% of ECG signals when generating explanations, meaning doctors can understand model decisions by examining smaller, more focused portions of heart recordings rather than entire 12-lead datasets—making explanations more actionable in time-constrained clinical settings.

Speed enables real-world deployment: The system generates explanations in under 1 second, which is fast enough for near real-time clinical use, addressing a practical barrier to implementing explainable AI in actual healthcare workflows where delays could impact patient care.

Performance varies significantly by condition: While the method achieves 98.9% validity for myocardial infarction (MI) detection, it struggles with hypertrophy (HYP) detection at only 13.2% validity, highlighting that explainability solutions may work better for some medical conditions than others—an important consideration before clinical deployment.

Physiological alignment matters for trust: By aligning explanations to R-peaks (natural cardiac landmarks), the framework generates counterfactuals that respect the heart's actual physiology rather than arbitrary signal modifications, which could increase clinician trust and adoption compared to purely statistical explanations.

Foundation for interactive systems: The research establishes design principles for building explanation platforms where clinicians could control and customize how AI decisions are explained to them, moving toward more collaborative human-AI diagnostic tools rather than one-size-fits-all explanations.