LICE-guided Sensitivity-oriented Model Refinement for Diabetes Prediction: An Ablation Analysis of Interaction Features and Sample Weighting

Diabetes prediction models support early risk identification, but false-negative predictions may leave diabetic cases undetected.This work presents a LICE-guided refinement framework for diabetes prediction.Local Interpretability with Counterfactual Explanations (LICE) examines local feature contributions, generates constrained counterfactual changes, and informs feature construction and targeted sample weighting.A tuned Gradient Boosting Decision Tree model served as the baseline for the BRFSS2015 dataset.We applied LICE to out-of-fold false-negative cases and compared eight predefined ablation configurations on a held-out test partition.Among the prespecified configurations, LICE-BalancedGBDT reduced false negatives by 172 cases, a 12.14 percent relative reduction, with recall of 0.8244, F1-score of 0.7770, MCC of 0.5163, and AUC of 0.8307.The findings show that LICE-guided refinement shifted prediction behaviour toward fewer false negatives, accompanied by a small numerical decrease in AUC relative to the baseline.