Beyond SHAP: Reliable feature selection methods for clinical prediction models

Abstract

This study critically examines the limitations of model-dependent feature importance methods used in clinical prediction modeling, specifically addressing inconsistencies in Xu et al.'s (2025) depression prediction research. We demonstrate how algorithm selection fundamentally alters featured rankings despite similar prediction accuracies, revealing a methodological gap where accuracy validation exists but feature importance validation does not. We propose a comprehensive alternative framework combining statistical and information-theoretic approaches: (1) monotonic relationship detection using Spearman's correlation and Kendall's tau with p-value assessment, and (2) complex interaction analysis using Mutual Information and Effective Transfer Entropy. This dual methodology enables identification of both straightforward variable associations and complex nonlinear dependencies, providing more robust and reliable insights for clinical prediction models.