RT2C: Predicting Time-to-New Caries with Structured Dental Data Using RNN
X Liu1,2, K K Kookal3, R Brandon4
1McWilliams School of Biomedical Informatics, University of Texas Health Science Center in Houston, Houston, TX, USA.
Journal of Dental Research
|July 18, 2026
Summary
A new recurrent neural network model, RT2C, accurately predicts dental caries risk and time-to-event, improving upon existing methods for personalized patient care.
Area of Science:
- Dentistry
- Machine Learning
- Public Health
Background:
- Dental caries is a prevalent chronic condition.
- Existing risk models often provide binary predictions and lack time-to-event data.
- Accurate caries risk assessment is crucial for preventive dental care.
Purpose of the Study:
- To develop and validate RT2C, a survival model using recurrent neural networks (RNNs).
- To assess the likelihood and time-to-event for new caries diagnosis.
- To support personalized caries management within the Caries Management by Risk Assessment (CAMBRA) framework.
Main Methods:
- Utilized structured longitudinal electronic dental record (EDR) data from 466,782 patients.
- Employed a bidirectional gated recurrent unit network (RT2C model).
- Compared RT2C against Cox proportional hazards and random survival forest models using concordance index and Brier score.
Main Results:
- RT2C achieved a high concordance index (0.883 ± 0.08) on the test set, outperforming baseline models.
- Model performance was consistent across diverse demographic and clinical subgroups, indicating fairness and robustness.
- RT2C effectively models visit-indexed risk trajectories and predicts new caries occurrence by the subsequent visit.
Conclusions:
- RT2C offers a valuable advancement for caries risk assessment and management.
- The model integrates longitudinal EDR data to provide personalized, proactive dental care.
- RT2C enhances existing Caries Risk Assessment (CRA) tools, aligning with the CAMBRA preventive framework.