CA-LIME: A Collinearity-Aware LIME Framework for Reliable and Interpretable Explanations in Healthcare AI

In clinical tabular models, correlated biomarkers can make local explanations unstable even when predictive performance appears acceptable. This paper presents CA-LIME, a collinearity-aware extension of LIME built to address that failure mode directly. Five variations of CA-LIME Vanilla reference, VIF, PCA back projection, Ridge, and ElasticNet allow us to investigate how various correlation-control strategies alter explanation behavior. ILPD, Kaggle liver, a combined ILPD + Kaggle liver cohort, and heart stroke are the four dataset configurations on which we test the system. Existing validated ILPD/stroke assets are reused, and the same workflow is extended to the new liver parameters for consistency. With an emphasis on local interpretability, stability, clinical plausibility, and practical application, we also contrast CA-LIME with SHAP as the baseline explanation. Across settings, CA-LIME retains clinically meaningful feature emphasis and shows stable local behavior under correlated inputs. The findings clarify the remaining validation and deployment constraints while supporting CA-LIME as a decision-support explanation layer.