Interpretable Machine Learning for Population-Level Tooth Loss Prediction

Machine learning can support population-level severe tooth loss (STL; ≥6 missing teeth) risk stratification; however, a lack of calibration under domain shift, limited interpretability of conventional black-box models, and inadequate handling of complex survey designs constrain responsible public health interpretation and implementation. We implemented and evaluated an interpretable, survey-weighted Multiple Imputation by Chained Equations-Explainable Boosting Machine (MICE-EBM) framework for population-level STL prediction incorporating sociobehavioral and systemic health determinants. Representative US adult datasets were analyzed, including Behavioral Risk Factor Surveillance System (BRFSS) 2022 for model derivation (N = 433,772), BRFSS 2024 for temporal validation (N = 448,213), and National Health and Nutrition Examination Survey (NHANES) 2015-2018 for cross-survey evaluation under outcome-definition and measurement shift (N = 10,775). Missing data were addressed using an antileakage HistGradientBoosting-driven pipeline to preserve multivariate epidemiological variance. The EBM demonstrated strong temporal stability on BRFSS 2024 (area under the receiver-operating characteristic curve [AUC]: 0.863; Brier: 0.085). For BRFSS 2022, model performance yielded an AUC of 0.865 and a Brier score of 0.086; a 100-replicate locked-imputed-set bootstrap confirmed robustness with an optimism-corrected AUC of 0.860 and a Brier score of 0.088. Direct transfer to NHANES yielded an AUC of 0.754 with poor raw calibration (Brier: 0.192), whereas predefined isotonic recalibration improved the holdout Brier score to 0.136. Compared with black-box stacked meta-ensemble (AUC: 0.780), the pre-recalibration EBM had lower NHANES discrimination (AUC: 0.754) but retained intrinsic glass-box interpretability. Overall, the MICE-EBM model may support population-level STL risk stratification for US survey-based cohorts with intrinsic transparency and calibrated same-survey performance. Moreover, this TRIPOD+AI-compliant framework may support population-level dental public health planning, although cross-survey or international application requires local validation and recalibration prior to implementation.