Predicting the prognosis of endodontic treatment in primary teeth through a machine learning approach

• Grant number: 25/14238-6
• Support Opportunities: Scholarships abroad - Research Internship - Master's degree
• Start date: January 01, 2026
• End date: March 31, 2026
• Field of knowledge: Health Sciences - Dentistry - Pediatric Dentistry
• Principal Investigator: Fausto Medeiros Mendes
• Grantee: Gustavo da Costa Fernandes
• Supervisor: Helena Silveira Schuch
• Host Institution: Faculdade de Odontologia (FO). Universidade de São Paulo (USP). São Paulo , SP, Brazil
• Institution abroad: University of Queensland, Brisbane (UQ), Australia
• Associated to the scholarship: 23/18210-3 - Evaluation of patient-centered outcomes after non-instrumentation endodontic treatment of primary molars - a multicenter randomized clinical trial with 24 months of follow-up, BP.MS

Abstract

Although advancements have been made in the endodontic treatment of primary teeth, literature on this topic remains limited compared to other dental procedures. The factors that influence the success of this treatment remain unclear. However, with the advent of artificial intelligence, models created using machine learning (ML) techniques can now predict the success of this type of treatment with greater accuracy. Nevertheless, no previous study has used an ML approach for this purpose. Therefore, the aim of this study is to develop ML algorithms to predict the success of primary teeth endodontically treated. For this purpose, data from five clinical trials conducted at our institution will be utilized. These trials included approximately 555 and 365 treated teeth, which were followed up for 12 or 24 months, respectively. The predictors will include child- (sex, age), teeth- (tooth type, arch, side, initial pulpal condition, presence of root resorption or periapical lesion), and endodontic procedures-related variables (instrumentation technique, filling, and restorative material used). The outcome will be clinical and radiographic failures after one year and after two years of treatment. For model training and development, the sample will be divided into a training and test set (70:30), and standard ML algorithms will be used. After optimizing hyperparameters, the area under the ROC curves will be derived, and other accuracy parameters will be calculated using the incidence of failures in the training sample as the cut-off point. Evaluation metrics will also include sensitivity, specificity, and Brier score. SHapley Additive exPlanations (SHAP) values will be plotted to identify the top predictors for each model.