Using Explainability as a Training-Time Reliability Signal for Efficient ECG Classification

Researchers introduce ERTS, an explainability-based training method that reduces computational costs for ECG classification by using attention map quality to identify which training samples are genuinely informative versus noisy. The approach demonstrates consistent performance improvements across multiple datasets while significantly lowering training expenses, offering practical efficiency gains for resource-constrained healthcare environments.

This research addresses a critical bottleneck in clinical machine learning: the prohibitive computational cost of training deep networks on medical time-series data. Healthcare institutions frequently lack the infrastructure for repeated model development cycles, making efficiency gains directly translatable to broader adoption of AI-driven diagnostic tools. ERTS tackles this by moving beyond simple confidence-based sample filtering to leverage explainability mechanisms as a reliability signal during training.

The innovation builds on progressive data dropout—a technique that accelerates training by excluding already-learned samples from gradient updates. However, traditional confidence metrics conflate genuine learning difficulty with noise or label ambiguity. ERTS distinguishes between these by computing Grad-CAM attention maps and deriving a focus score that validates whether model predictions align with coherent, localized patterns in ECG data. Samples with weak attention focus are filtered, while those with meaningful spatial-temporal patterns receive priority.

From a clinical AI perspective, this represents meaningful progress toward democratizing model development. The consistent macro-F1 improvements across three ECG datasets and multiple architectures suggest the approach generalizes beyond specific implementations. For healthcare providers and researchers operating under budget constraints, reduced training costs directly enable more frequent model updates, better dataset diversity exploration, and faster iteration cycles.

The intersection of explainability and efficiency proves particularly valuable in regulated domains where interpretability increasingly influences deployment decisions. As healthcare systems face pressure to adopt AI while maintaining transparency and resource efficiency, methods that simultaneously improve both reliability signals and computational efficiency gain substantial relevance in vendor selection and internal development priorities.

• →ERTS uses Grad-CAM attention map quality as a training signal to distinguish informative samples from noisy ambiguous ones in ECG classification.
• →The method achieves consistent macro-F1 improvements while reducing effective training costs across three datasets and multiple model architectures.
• →Explainability-based filtering proves more effective than confidence-based sample selection for clinical time-series learning.
• →The approach makes deep learning model development more feasible for resource-constrained healthcare settings.
• →Results suggest explanation quality serves dual purposes: improving both computational efficiency and clinical reliability simultaneously.