External Validation of a Machine Learning Model for Oral Mucositis Prediction In Proton Therapy
Abstract
Purpose
Proton therapy’s ability to spare normal tissue generally lowers complications, making it an attractive treatment option for head and neck (HN) cancer. Yet, oral mucositis, one of the most frequent and clinically relevant toxicities, occurs in approximately 60% of patients. This clinical outcome underscores the limited effectiveness of current clinical toxicity mitigation strategies, which primarily rely on one-dimensional dose constraints derived from photon therapy. We recently developed a machine learning (ML) risk model to predict oral mucositis in HN patients. Here, we will present the updated model and its external validation. ML allows the creation of a multidimensional model, beyond dose, which improves prediction accuracy compared to standard normal tissue complication probability (NTCP) models.
Methods
Our random forest ML was trained and tested on 87 HN patients treated with proton therapy at the University of Miami (UM). The model uses 21 features selected among dose-volume histograms (DVHs) and dose-yD-volume histograms (DyVHs), which integrate yD, a microdosimetric surrogate of LETD capturing energy deposition at the microscopic scale, into the DVH. The external validation cohort includes 24 patients treated at the South Florida Proton Therapy Institute (SFPTI). DVHs were extracted from the clinical treatment plan; yD values were calculated with OpenTOPAS.
Results
The ML model achieved a ROC–AUC of 0.85 for the UM cohort and 0.79 for the SFPTI cohort, always outperforming the NTCP approach.
Conclusion
The ML model demonstrates robust predictive performance in both internal and external cohorts. Additional external validation using a cohort from the New York Proton Center is ongoing and will further expand tumor site and population diversity. Ultimately, we aim to integrate this model into the RayStation treatment planning system to enable toxicity-aware plan optimization and reduce oral mucositis without compromising target coverage.