Dental services use prediction among adults in Southern Brazil: A gender and racial fairness-oriented machine learning approach

Objective

To develop machine learning models to predict the use of dental services among adults aged 18 and older.

Methods

This is a prospective cohort study that uses data from the survey “EAI Pelotas?”. The sample consisted of individuals who participated in both the baseline and follow-up, totaling 3461 people. Predictors were collected as baseline and comprised 47 sociodemographic, behavioral, oral and general health characteristics. The outcome was dental service use in the last year assessed during the one-year follow-up. Data was divided into training (80 %) and test (20 %) sets. Five machine learning models were tested. Hyperparameter tuning was optimized through 10-fold cross-validation, utilizing 30 iterations. Model performance was assessed based on the area under the Receiver Operating Characteristic (ROC) curve (AUC), accuracy, recall, precision, and F1-score.

Results

The prevalence of dental service use in the follow-up was 47.2 % (95 % CI, 45.5 – 48.9). All models in the test set demonstrated an AUC-ROC between 0.76 and 0.77. The CatBoost Classifier model exhibited the highest performance in the test dataset among the models concerning the AUC metric (AUC = 0.77, CI95 %,[0.73–0.80]), displaying an accuracy = 0.69, recall = 0.69, precision = 0.68, and F1-score = 0.69. Fairness estimations for the best model indicated consistent performance across gender categories. However, disparities were observed among racial groups, AUC = 0.57 for individuals who self-reported mixed (“pardos”) skin color. The explainability analysis shows that the most important features were the last dental visit at baseline and education level.

Conclusion

Despite our findings suggesting a sufficient prediction of overall dental services’ use, performance varied across racial groups.

Clinical significance

Our findings highlight the potential of machine learning models to predict dental service use with good overall accuracy. However, the significantly lower performance for mixed-race individuals raises concerns about fairness and equity. Therefore, despite promising results, the model requires further refinement before it can be applied in real-world public health settings.