Interpretable Lifestyle-Only Diabetes Screening: A Clinically Grounded XAI Pipeline

Abstract

This paper presents an explainable AI pipeline for diabetes screening that learns from non-laboratory, self-reported health indicators and then provides patient-specific, actionable risk rationales clinicians can audit and communicate. A deep neural classifier, gradient boosting, and a regularized logistic baseline are trained on a large BRFSS-derived tabular cohort with severe class imbalance; the logistic baseline provides stable feature attributions for lifestyle factors while the black-box models optimize discrimination, reflecting best practices in XAI for healthcare tabular data. The system attains verified discrimination between diabetes-free and diabetes cases on the held-out test set, with macro-averaged F1-scores of 0.432 (DNN), 0.423 (XGBoost), and 0.389 (logistic regression), computed with macro averaging over the three classes; these values are consistent with the class-level F1-scores reported and with the underlying held-out test predictions. Explanations are based on regularized logistic regression coefficients aligned with established XAI methods, and the system is positioned as preliminary risk stratification rather than clinically validated screening. The pre-diabetes class remains poorly identified (F1  ≤ 0.050 for all models), and claims of equitable early intervention in resource-limited settings require further prospective validation.