Development of a Machine Learning Tool for Home-Based Assessment of Periodontitis

According to an ADA report, approximately 15% of the US population requires dental care annually but does not receive it. Access to dental care, particularly for periodontal examinations, is challenging for many individuals, leading to uncontrolled periodontitis progression and systemic health complications. Periodontitis, an inflammatory gum disease, affects nearly half of American adults over 30. Current diagnostic approaches rely on periodontal exams and radiographs, requiring clinical settings and experienced dental care providers. However, many individuals lack access to dental care, making it difficult to obtain up-to-date clinical probing depth, dental X-rays or CT scans. To address this gap, we developed a machine learning (ML) tool for at-home preliminary periodontitis assessments. This tool would benefit individuals unaware of their undiagnosed periodontal conditions and those with limited access to dental care, empowering them to prioritize dental care and seek timely treatment within their constraints. Our tool leverages the NHANES database to train an ML model on multimodal features relevant to periodontitis that are radiographic-independent. We labeled the individuals with different periodontitis severity based on their periodontal charting records and performed feature engineering on the dataset. We first developed a baseline model and subsequently trained additional classifiers, conducting a comprehensive hyperparameter search that resulted in consistent performance. The best-performing model was evaluated on the test set, achieving an overall precision of 0.80 and AUC of 0.81, demonstrating robust classification performance without overfitting. Feature importance analysis provided guidance for the questionnaire design for the real-world application of this tool. Additionally, our novel approach of analyzing misclassified populations offered insights for data interpretation, supported model improvement, and revealed deeper correlations between periodontitis and its risk factors. Our model exemplifies the capacity to leverage extensive public health databases for periodontitis evaluations. Ultimately, our ML-driven tool aims to overcome existing dental care barriers by providing users with periodontitis predictions and personalized dental care suggestions, all easily accessible from their smartphones or laptops at home. One sentence summaryIn this study, we developed the machine-learning (ML) model for home-based screening of periodontitis and demonstrated that non-radiographic data--encompassing demographic and nutrition information, as well as medical and oral health conditions--possesses strong predictive power for periodontitis, empowering individuals with limited access to dental care to properly identify their periodontal health status and promote timely intervention.