Distributional reliability explanation for diffusion-based electrocardiogram abnormality risk prediction

Abstract

Probabilistic machine-learning models can represent predictive uncertainty, but conventional explanation methods usually summarize only point-output behavior. This is limiting in risk-prediction settings where uncertainty, tail probability, and the structure of the predictive distribution are central to model reliability. In this study, we applied distributional reliability explanation attribution (DREA) to examine how electrocardiogram-derived predictors influence the full risk distribution generated by a diffusion-based abnormality-risk model, in comparison with conventional point-output explanations. We used the PTB-XL electrocardiogram dataset and constructed lead-specific engineered, physiological, and waveform-derived descriptors. On the held-out patient-disjoint test set, the diffusion predictive mean achieved an area under the receiver operating characteristic curve of 0.858 and an area under the precision-recall curve of 0.876, compared with 0.870 and 0.878, respectively, for the Random-Forest classifier. Although the diffusion model was not the strongest point predictor, its probabilistic output enabled distribution-level analysis of model reliability. Point-output and distributional explanations captured complementary forms of model behavior. Age was highly important under both point-output SHAP analyses and retained high DREA importance, whereas Lead V1 waveform-change intensity showed lower Random-Forest SHAP importance but high diffusion-mean SHAP and the highest mean DREA score. Across 80 independently retrained model realizations, the leading DREA rankings were stable and controlled what-if perturbations provided complementary evidence that highly ranked predictors produced comparatively large scenario-specific changes in distributional shift, high-risk exceedance probability, mean risk, predictive width, and variance. These findings position DREA as a complementary reliability-auditing layer for probabilistic risk models, particularly when predictive width, variance, tail behavior, and the broader structure of the predictive distribution matter alongside discrimination performance.