Machine learning‐assisted prediction of clinical responses to periodontal treatment
Balázs Fehér, Eduardo Henrique de Souza OLIVEIRA, Poliana Mendes Duarte, Andreas A. Werdich, William V. Giannobile, Magda Feres
Abstract
Abstract Background Periodontitis is among the most prevalent chronic inflammatory conditions globally, and is associated with bone resorption, tooth loss, and systemic complications. While its treatment is largely standardized, individual outcomes vary, with some patients experiencing further disease progression despite adherence. Methods We developed a machine learning (ML) approach to predict individual outcomes 1 year post‐treatment using retrospectively assessed baseline parameters. We trained a Random Forest model on 18 demographic, clinical, microbiological, and treatment‐related features of 414 patients from randomized clinical trials (RCTs) in South America. We subsequently performed internal testing, interpretability analysis, and external testing on a second dataset of 78 patients from previous RCTs in North America and Europe exhibiting less severe disease. Results In internal testing, the ML model achieved an area under the receiver operator characteristics curve (AUROC) of 0.93, an area under the precision‐recall curve (AUPRC) of 0.90, an F 1 ‐score of 0.82, and an out‐of‐bag score of 0.71. Relative importances were 0.42 for clinical, 0.33 for treatment‐related, 0.21 for microbiological, and 0.04 for demographic features. In external testing, the ML model achieved an AUROC of 0.76, an AUPRC of 0.69, and an F 1 ‐score of 0.71. Conclusions Our study indicates that an ML‐based approach can assist in predicting individual responses to periodontal treatment. Prospective validation is needed for clinical application. Plain language summary Using comprehensive data from patients with periodontitis, an inflammatory condition of the tooth‐supporting tissues, a machine learning model was trained to predict how well patients might respond to different treatments after 1 year. The model was externally tested in patient populations from 2 continents different from the training dataset. The results suggest that with further research and refinement, this tool could eventually become a valuable asset in personalizing treatment plans for improved patient outcomes.