Multi-institutional Normal Tissue Complication Probability (NTCP) Prediction Model for Mandibular Osteoradionecrosis:...

Mandibular osteoradionecrosis (ORN) is a severe late complication affecting patients with head and neck cancer (HNC) treated with radiation therapy (RT) that significantly impacts patients' quality of life and can require costly interventions. Although radiation dose is a key factor, other clinical and demographic risk factors also influence ORN development. Previous predictive models have primarily been single-institutional, limiting their generalizability. In this first analysis from the PREDMORN Consortium, we have aimed to reproduce existing statistical association and modeling analyses on the largest and most diverse mandibular ORN cohort worldwide to allow comparison with previous studies.

This retrospective multi-institutional study included 3928 patients with HNC (622 ORN cases) from 8 institutions. Clinical, demographic, and dosimetric variables were analyzed to develop a prediction model (any grade ORN vs no ORN) using forward stepwise logistic regression with correlation-based variable preselection. The ORN normal tissue complication probability (NTCP) model was developed on 80% of data from 6 institutions, tested on the remaining unseen 20%, and externally validated on a matched cohort (58 patients, 19 ORN cases) and a large population-based cohort (2687 patients, 215 ORN cases).

Key predictors of ORN were D_30%, V_70Gy, pre-RT dental extractions, and smoking status. The ORN NTCP model demonstrated very good calibration on the population-based external cohort (Brier score, 0.077; Log Loss, 0.281). Model discrimination improved on a subcohort including oropharyngeal and locally advanced larynx/hypopharynx cancer cases only (AUC from 0.69 to 0.75 and from 0.65 to 0.67 on the matched and the population-based external cohorts, respectively).

The PREDMORN NTCP model is the largest multi-institutional effort to date aimed at predicting ORN risk in patients with HNC using real-world data. The model demonstrated good generalizability when externally validated to a large population-based cohort. Our observations align with current guidelines and corroborate findings from smaller single-institution studies.