Work

107. Artificial Intelligence-Enabled Prediction of Osteoradionecrosis Following Transferred Fibula Flap for Head and Neck Tumor Resection and Segmental Mandibulectomy

Year: 2025

Type: article

Abstract: PURPOSE: Osteoradionecrosis (ORN) is a severe complication that can occur following fibular free flap reconstruction for tumor resection and segmental mandibulectomy, potentially leading to significant morbidity and reduced quality of life. Accurate prediction of ORN risk is essential for optimizing postoperative planning and improving patient outcomes. However, predicting ORN is challenging due to the complexity of patient factors, tumor characteristics, and diverse treatment modalities. Currently, robust algorithms for ORN risk prediction are lacking. This study aimed to evaluate the effectiveness of a traditional nomogram and seven machine learning (ML) algorithms in predicting ORN. METHODS: In this retrospective cohort study, all patients who underwent transferred fibula flap for head and neck tumor resection and segmental mandibulectomy at Chang Gung Memorial Hospital, Taiwan, from 2014-2019 were identified (n=333). After excluding 59 patients due to incomplete data, 274 patients were included. Patients were randomly divided into training (n=178) and testing (n=96) sets in a 65:35 ratio. A nomogram and seven ML algorithms (Decision Tree, K-Nearest Neighbors, Neural Network, Random Forest, Support Vector Machine, Voting Ensemble, and XGBoost) were trained and evaluated in predicting ORN. Model assessment metrics included Area Under the Receiver Operating Characteristics (AUROC), accuracy, precision, specificity, sensitivity, and F1-score. SHapley Additive exPlanations (SHAP) evaluated feature importance, while Decision Curve Analysis (DCA) assessed clinical utility. RESULTS: The Neural Network model showed the highest predictive performance (AUROC: 0.91, 95% CI: 0.84-0.97), significantly outperforming the nomogram (AUROC: 0.86, 95% CI: 0.76-0.93; p=0.022). The Vote model also surpassed the nomogram (AUROC: 0.91, 95% CI: 0.84-0.96; p=0.042). SHAP analysis identified wound infection, plate exposure, post-op radiotherapy, and tumor stage as key predictive factors across all models. The nomogram uniquely emphasized post-operative albumin; the Neural Network focused on number of intra-operative fibula osteotomies; the Vote model highlighted tumor site and Jewer-Boyd mandibular defect classification. Of the 75 non-ORN and 21 ORN cases in the testing set, the nomogram correctly predicted 72 non-ORN and 8 ORN cases; the Neural Network identified 66 non-ORN and 17 ORN cases; the Vote model predicted 64 non-ORN and 18 ORN cases. The DCA showed that the Neural Network model provided the highest net benefit when the risk of ORN was low (0-20% ORN risk), while the Vote model was optimal for medium risk levels (20-60% ORN risk). CONCLUSION: The Neural Network and Voting Ensemble models exhibited excellent predictive accuracy for ORN, with AUROC scores exceeding 0.90, significantly outperforming a traditional nomogram. SHAP analysis ensured the interpretability of these models by identifying key predictive factors specific to each algorithm. Additionally, decision curve analysis (DCA) demonstrated the high clinical utility of both models across varying ORN risk thresholds. By combining enhanced accuracy, interpretability, and clinical relevance, these ML models represent a significant advancement in ORN risk assessment, potentially leading to more personalized treatment approaches and improved patient outcomes in head and neck reconstruction.

Source: Plastic & Reconstructive Surgery Global Open

DOI: 10.1097/01.gox.0001112376.20471.eb

Authors: Dany Y. Matar, Gina Mackert, Anthony Y. Matar, Angela Chien-Yu Chen, Sashank Reddy, Huang-Kai Kao