Instance level analysis of equitable learning via dissimilar variable grouping on healthcare datasets.

Abstract

Machine learning models are increasingly deployed in healthcare, but their complexity can propagate biases and obscure decision-making that disproportionately affect vulnerable populations. While existing explainability methods focus on variable-level attribution, they often fail to capture how predictions and explanations behave at the instance level. To address this gap, we introduce SHIELD: a SHapley and Information-theory based framework for Equitable Learning via Dissimilar variable grouping. In this framework, variables are grouped using conditional mutual information to weaken correlations that may encode sensitive attributes. Group-specific autoencoders learn latent representations that preserve a mapping to the original variables, allowing SHapley Additive exPlanations (SHAP) to be computed in the original variable space for faithful instance-level attribution. Experiments on three clinical datasets demonstrated that dissimilarity-based grouping produced a more even distribution of variable importance and increased total attribution magnitude, suggesting broader usage of variables in model decision-making. On average, grouping improved the six evaluated fairness measures, covering both outcome and explanation disparities by 17.09%. Nevertheless, predictive performance decreased by 6.18% across six metrics. These changes represent a context-dependent trade-off rather than directly comparable gains and losses. Overall, SHIELD provides a principled and reproducible framework for equitable and explainable machine learning in health informatics, with code available to support future research.