A selective machine learning algorithm for severe periodontitis labeling from questionnaire data.

Area of Science:

• Oral health research
• Epidemiology
• Machine learning applications

Background:

• Epidemiological studies frequently collect self-reported oral health (SROH) data but lack clinical periodontal measurements.
• Accurate identification of periodontitis cases is crucial for cohort studies but challenging without clinical examinations.

Purpose of the Study:

• To develop and validate a machine learning (ML) pipeline for selective identification of severe periodontitis (SP) and no periodontitis (NP) from SROH questionnaires.
• To support case-control selection in large cohorts where clinical periodontal data is unavailable.

Main Methods:

• A two-stage ML pipeline using CatBoost was developed on three datasets (n=498) with SROH, demographics, and CPITN scores.
• The pipeline involved data cleaning, feature engineering, and a two-model approach (Separator-A and Separator-Z) with high-confidence prediction thresholds (≥0.85).
• Model performance was evaluated on internal test and hold-out inference sets, with a focus on accurate classification of NP and SP cases while excluding moderate periodontitis (MP).

Main Results:

• The ML pipeline achieved complete separation between NP and SP cases within a high-confidence subset (4.31% of eligible cases).
• No moderate periodontitis (MP) cases were misclassified as NP or SP.
• The pipeline demonstrated high performance and explainability in selectively identifying periodontitis status from questionnaire data.

Conclusions:

• A selective, explainable two-stage ML pipeline can accurately identify severe periodontitis and no periodontitis from SROH questionnaire data.
• This approach offers a valuable tool for case-control selection in epidemiological cohorts lacking clinical periodontal examinations.
• Further validation is recommended to confirm the generalizability of the pipeline across diverse populations.