Suitability of machine learning models for prediction of clinically defined Stage III/IV periodontitis from questionnaires and demographic data in Danish cohorts

Aim: To evaluate if, and to what extent, machine learning models can capture clinically defined Stage III/IV periodontitis from self-report questionnaires and demographic data. Materials and Methods: Self-reported measures of periodontitis, demographic data and clinically established Stage III/IV periodontitis status were extracted from two Danish population-based cohorts (The Copenhagen Aging and Midlife Biobank [CAMB] and The Danish Health Examination Survey [DANHES]) and used to develop cross-validated machine learning models for the prediction of clinically established Stage III/IV periodontitis. Models were trained using 10-fold cross-validations repeated three times on the CAMB dataset (n = 1476), and the resulting models were validated in the DANHES dataset (n = 3585). Results: The prevalence of Stage III/IV periodontitis was 23.2% (n = 342) in the CAMB dataset and 9.3% (n = 335) in the DANHES dataset. For the prediction of clinically established Stage III/IV periodontitis in the CAMB cohort, models reached area under the receiver operating characteristics (AUROCs) of 0.67–0.69, sensitivities of 0.58–0.64 and specificities of 0.71–0.80. In the DANHES cohort, models derived from the CAMB cohort achieved AUROCs of 0.64–0.70, sensitivities of 0.44–0.63 and specificities of 0.75–0.84. Conclusions: Applying cross-validated machine learning algorithms to demographic data and self-reported measures of periodontitis resulted in models with modest capabilities for the prediction of Stage III/IV periodontitis in two Danish cohorts.