Conference Agenda

The increasing use of predictive models in clinical medicine underscores the need for interpretability mechanisms capable of elucidating the internal logic behind algorithmic decisions. Although black box models often achieve high predictive accuracy, their limited transparency poses significant barriers to clinical adoption. This study presents a comprehensive in silico evaluation framework for Explainable Artificial Intelligence (XAI) techniques, focusing on SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model Agnostic Explanations) applied to supervised models trained on synthetically generated clinical cohorts. A reproducible benchmarking methodology was developed to compare XAI approaches across multiple model architectures (Decision Trees, Random Forest, XGBoost) under carefully controlled simulation scenarios. Technical performance metrics—including fidelity, consistency, and stability—were computed across Monte Carlo replicates with bootstrap derived confidence intervals. SHAP demonstrated superior fidelity (0.89 ± 0.04, 95% CI [0.85, 0.93]) and stability (0.81 ± 0.07) relative to LIME, which achieved a fidelity of 0.74 ± 0.08. Robustness analyses showed that explanation stability decreased by 15.2% (p < 0.001) under 5% input perturbations, with substantial variability across noise levels (η² = 0.34).The proposed framework provides evidence based guidance for selecting and validating XAI methods in clinical decision support applications, emphasizing reproducibility and methodological rigor in alignment with emerging governance standards for AI transparency.