Machine Learning Models for Identifying Dental Pain in Adolescents

Introduction: This study aimed to identify dental pain using machine learning (ML) algorithms in Brazilian adolescents for public health screening purposes. Methods: Data from 2 cross-sectional waves of the Brazilian National Survey of School Health (PeNSE) in 2015 and 2019 were used (schoolchildren aged 11 to 18). The outcome was dental pain in the last 6 months. Co-variables were 53 variables, including demographic, socioeconomic, and behavioral characteristics. The 2015 dataset was split (80:20) into training and test sets, while the 2019 dataset was used as a temporal external validation set. Nine ML models were evaluated. Results: A total of 259,833 adolescents (97.0% of the sample) were included. Dental pain prevalence was 19.5% (95% CI, 19.2-19.8). Extra Trees (ET) was the model with the best metrics in the test and external validation sets. ET showed an AUC = 0.64 (95% CI, 0.63-0.65) and a Recall = 0.57 in the test, and AUC = 0.62 (95% CI, 0.62-0.63) and Recall = 0.57 in the external test, indicating a modest ability to discriminate adolescents with dental pain and to identify approximately 57 out of 100 affected individuals. Fairness estimations show lower accuracy for males, but a higher recall for this group. The model shows a higher accuracy for white adolescents but a lower recall for this group. The Shapley values showed that sex, alcohol consumption, and family violence were the most important variables in the algorithm's identification process. Conclusion: This study shows the potential of ML to identify dental pain in adolescents. Modest predictive performance and fairness limitations highlight the need for improvements before widespread adoption.