HybridXAI-Disease: A SHAP-Driven Framework for Actionable and Clinically Interpretable Disease Prediction
Year: 2026
Type: conference-paper
Abstract: Applying machine learning to clinical diagnostics requires explainable and interpretable decision-support systems. This paper presents HybridXAI-Disease, a framework that combines SHAP (SHapley Additive exPlanations) analysis, a SHAP-derived salience visualization, and bounded counterfactual reasoning into a single pipeline for disease prediction. Unlike prior work that relies on a single explainability technique, the framework provides three complementary explanation layers: global feature importance (SHAP), patient-level attribution (SHAP waterfall plots and cohort-wise salience maps), and actionable, minimum-change counterfactual guidance. Five classifiers were trained and compared on the Wisconsin Breast Cancer Dataset (WBCD, n=569); the highest test-set accuracy was 97.37% (SVM and Random Forest) with AUC-ROC up to 99.60% (Logistic Regression). XGBoost, the base classifier for the explanation pipeline, achieved 96.49% accuracy and 99.40% AUC-ROC, a small trade-off made explicit and justified by its exact, polynomial-time SHAP attribution. The morphological features worst concave points, area error, and mean concave points were the dominant predictors of malignancy. It also provides a dedicated ablation study – validating the salience layer against permutation importance (Spearman ρ=0.53, p=0.0025), justifying the counterfactual feature budget K=5 via a Top-K sufficiency curve, as well as directly measuring counterfactual flip rate (84.2%) and plausibility (53.1% of successful flips remain within the training distribution without clipping) – alongside a calibration analysis (Brier scores, reliability diagrams) and explicit discussion of practical deployment limitations.
Institution: SR University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7224970641