JMIR Publications

ABSTRACT

Background:

Chronic periodontitis affects 26% in Thai adults and 11.2% globally, resulting in loss of tooth and quality of life. Although periodontal probing is the gold standard, it is time and resource-consuming, so a screening model to identify chronic periodontitis can aid in reducing workload for dentists. While cross-sectional logistic regression is common to apply, optimal performance depends on feature selection and engineering. Machine learning recently has been applied due to their complex yet powerful performances.

Objective:

We aim to compare the performance of screening models developed on statistical and machine-learning principles.

Methods:

This study used data from the prospective Electricity Generating Authority of Thailand (EGAT) cohort. Dental examinations were performed for 2008 and 2013 surveys. The outcome of interest was periodontitis diagnosed by the Centre for Disease Control – American Academy of Periodontology defined criteria. Risk prediction models were developed using mixed-effects logistic regression (MELR), recurrent neural networks (RNN), mixed-effects support vector machine (ME-SVM), and mixed-effects decision tree (ME-DT).

Results:

Of 2,086 subjects, data were split into 80% and 20% for development and validation sets, respectively. As a result, 1,759 and 327 subjects were used for development and validation, with a prevalence of periodontitis of 34.4% and 34.1%, respectively. MELR (AUC and F-Score of 98.0% and 86.9%) performed better than RNN, ME-SVM and ME-DT (74.7% and 57.3%; 76.1% and 56.4%; 69.5% and 50.9% respectively) in identifying severe periodontitis.

Conclusions:

MELR model might be potentially applied as a screening model to evaluate the need for further periodontal evaluation. However, the model requires further external validation.