Abstract
Epidemiological cohorts often collect self-reported oral health (SROH) questionnaires but lack clinical periodontal measurements. We developed a selective, explainable machine learning (ML) pipeline that can assign labels for severe periodontitis (SP) or no periodontitis (NP). Three datasets (n = 498) with SROH questionnaires, demographics, and Community Periodontal Index of Treatment Needs (CPITN) scores were used to derive NP, moderate periodontitis (MP), and SP categories. MP cases were excluded from model development. After cleaning and feature engineering, non-similar label duplicates were removed. A CatBoost model (Separator-A) was trained with tenfold cross-validation; NP/SP predictions were retained when probability ≥ 0.85. From these outputs and domain rules, a rule-consistent subset was created to train a second model (Separator-Z). Performance was evaluated on internal test and hold-out inference sets. Next, the pipeline was applied to MP cases. The pipeline achieved complete separation across all evaluation sets within the retained high-confidence subset, which represented 4.31% of eligible NP/SP cases, while no MP cases were misclassified as NP or SP. Thus, a two-stage, explainable ML pipeline can selectively identify SP and NP from SROH questionnaire data, supporting case-control selection in cohorts without clinical periodontal examinations, though validation is warranted to confirm generalizability.