Abstract / Summary
Progression after non-segmental resection for osteoradionecrosis of the jaw (ORNJ) affects a substantial proportion of patients, and preoperative assessment cannot currently identify those at risk. We developed machine learning models based on incremental-region CT radiomics to predict progression after non-segmental resection. Patients with ORNJ who underwent non-segmental jaw resection between January 2014 and December 2024 ( n = 120) were divided chronologically into a training set ( n = 90) and an internal hold-out test set ( n = 30). Three incremental-region models were developed from preoperative CT images: R Seq (sequestrum), R Seq+BM (plus a perilesional bony margin) and R Seq+BM+ST (plus a perilesional bony margin and adjacent soft tissue). Six algorithms were compared for each configuration by repeated 10-fold cross-validation in the training set. Because the test set was then used to compare configurations, both the model with the highest test-set AUC and the cross-validation–selected model of the same configuration are reported. Performance was assessed by discrimination, calibration and decision curve analysis. Among 18 model–algorithm configurations, R Seq+BM+ST with XGBoost achieved the highest test-set discrimination (AUC, 0.875; 95% CI, 0.727–1.000; Brier score, 0.197). Because this model was selected on test-set performance, the cross-validation–selected model of the same configuration is reported alongside it as a conservative estimate (GBM; cross-validated AUC, 0.779; test AUC, 0.806). The test AUC of the primary model exceeded its own cross-validated AUC by 0.121, and the conservative model’s test AUC fell 0.069 below it. The two osseous configurations showed lower test-set discrimination (maximum test AUC, 0.764 for and 0.792 for ), but the three configurations were not clearly distinguished under cross-validation (best mean CV AUC, 0.766, 0.799, and 0.779). The primary model showed preliminary rule-out potential within the internal test cohort. Under cross-validation, the three-region configuration did not demonstrate a clear advantage over the two osseous configurations, with an incremental gain observed only in the test-set comparison. Alongside a conservative benchmark, these findings should be regarded as hypothesis-generating and warrant external multicenter validation before clinical consideration.