Artificial Intelligence Outperforms a Nomogram for Osteoradionecrosis Prognostication Following Fibula Free Flap Reconstruction in Oral Cancer Patients.

Background: Osteoradionecrosis (ORN) is a serious complication in oral cancer patients undergoing tumor excision and mandibular reconstruction, particularly after radiotherapy. This study compared a logistic regression-based nomogram with five machine learning (ML) algorithms to identify the most clinically useful ORN prognostication tool.

Methods: We retrospectively analyzed 275 irradiated oral cancer patients who underwent segmental mandibulectomy and immediate fibula flap reconstruction at a single center. Patients were split 75:25 into training and test cohorts. 34 patient variables were used to train a nomogram and five ML models (DNN, KNN, SVC, LightGBM, Stacked Ensemble). The primary outcome was ORN within 5-10 years. Binary prediction performance on the test cohort was assessed using AUROC, Calibration and Decision Curve Analysis (DCA). Feature contribution on model prediction was assessed.

Results: The Stacked model achieved the highest test AUROC (0.83, 95% CI: 0.70-0.94), outperforming the Nomogram (0.73, 95% CI: 0.57-0.86; p = 0.04), KNN (0.81), DNN (0.79), LightGBM (0.78), and SVC (0.74). DNN showed the best calibration (ICI: 0.07), followed by KNN (0.09), SVC (0.11), and the Nomogram (0.22). Pre-operative radiation therapy, post-operative wound infection, plate exposure, and surgical re-exploration were the most influential features in model predictions. DCA showed that DNN, Stacked, and SVC models provided the greatest net clinical benefit across decision thresholds.

Conclusions: ML models outperformed the Nomogram in predicting ORN. They showed strong binary discrimination and effective risk stratification. These findings support use of ML models for guiding postoperative surveillance and personalized care, warranting further validation.