Related Experiment Video
Updated: Jul 16, 2025
Oral Health Assessment by Lay Personnel for Older Adults
Published on: February 2, 2020
Development of a machine learning multiclass screening tool for periodontal health status based on non-clinical
Ke Deng1, Francesco Zonta2,3, Huan Yang3
1Shanghai PerioImplant Innovation Center, Department of Oral and Maxillofacial Implantology, National Clinical Research Center of Stomatology, Ninth People's Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China.
Aim:
To develop a multiclass non-clinical screening tool for periodontal disease and assess its accuracy for differentiating periodontal health, gingivitis and different stages of periodontitis.
Materials And Methods:
A cross-sectional diagnostic study on a convenience sample of 408 consecutive subjects was conducted by applying three non-clinical index tests estimating different features of the periodontal health-disease spectrum: a self-administered questionnaire, an oral rinse activated matrix metalloproteinase-8 (aMMP-8) point-of-care test (POCT) and determination of gingival bleeding on brushing (GBoB). Full-mouth periodontal examination was the reference standard. The periodontal diagnosis was made on the basis of the 2017 classification of periodontal diseases and conditions. Logistic regression and random forest (RF) analyses were performed to predict various periodontal diagnoses, and the accuracy measures were assessed.
Results:
Four-hundred and eight subjects were enrolled in this study, including those with periodontal health (16.2%), gingivitis (15.2%) and stage I (15.9%), stage II (15.9%), stage III (29.7%) and stage IV (7.1%) periodontitis. Nine predictors, namely 'gum disease' (Q1), 'a rating of gum/teeth health' (Q2), 'tooth cleaning' (Q3a), the symptom of 'loose teeth' (Q4), 'use of floss' (Q7), aMMP-8 POCT, self-reported GBoB, haemoglobin and age, resulted in high levels of accuracy in the RF classifier. High accuracy (area under the ROC curve > 0.94) was observed for the discrimination of three (health, gingivitis and periodontitis) and six classes (health, gingivitis, stages I, II, III and IV periodontitis). Confusion matrices showed that the misclassification of a periodontitis case as health or gingivitis was less than 1%-2%.
Conclusions:
Machine learning-based classifiers, such as RF analyses, are promising tools for multiclass assessment of periodontal health and disease in a non-clinical setting. Results need to be externally validated in appropriately sized independent samples (ClinicalTrials.gov NCT03928080).
