Clustered Partial Explanation: A Multilevel Framework for Contextual Model Interpretation in Clinical Decision Support

Abstract

The increasing integration of artificial intelligence into clinical decision‐making underscores the need for transparent and trustworthy predictive models. Explainable artificial intelligence has emerged to address this challenge; however, most existing approaches rely exclusively on global or local analyses, which fail to capture the heterogeneity of model behavior across patient populations. In this paper, we introduce the clustered partial explanation (CPE) framework, a novel three‐level methodology comprising global, partial, and local analyses, designed to bridge this gap. The proposed framework combines cluster‐based subgrouping with explainable artificial intelligence techniques to uncover latent variations in feature influence and model behavior. We apply the proposed framework to a breast cancer survival prediction task using a Random Forest model trained on the SEER dataset. The resulting partial explanations reveal heterogeneous feature importance patterns across clusters, highlighting interactions that remain obscured under global analysis. By validating global explanations at the subgroup level, the proposed framework enhances both interpretability and reliability, thereby supporting context‐aware clinical decision‐making. Finally, we discuss the implications of this additional interpretability layer for future research.

Ähnliche Arbeiten