PERSONALIZED MACHINE LEARNING-BASED RISK PREDICTION OF DENTAL DISEASES IN OLDER PATIENTS

Abstract

The review systematises machine learning models developed for the prediction of dental outcomes in older patients and appraises their methodological quality. Reported performance is shown to depend on study design more than on the choice of algorithm. Values range from an AUC of 0.71 to 0.99, and the highest of them belong to cross-sectional studies in which a prognostic task has been replaced by a diagnostic one. Moving to an independent sample lowers performance by 0.13–0.17 AUC units, whereas logistic regression that properly accounts for the structure of the data performs no worse than neural architectures. The specific features of the older patient are only partly represented in the models. Polypharmacy, hyposalivation, oral frailty and social isolation enter them fragmentarily, and the competing risk of death is almost never modelled. Conditions for clinical applicability are formulated, namely a longitudinal design with a predefined prediction horizon, external validation, assessment of calibration and reporting in line with TRIPOD+AI.