2026 Proffered Presentations
S269: INSTITUTION-SPECIFIC MACHINE LEARNING ALGORITHM FOR PREDICTING FACIAL NERVE OUTCOMES IN PATIENTS UNDERGOING MICROSURGERY FOR RESECTION OF VESTIBULAR SCHWANNOMA
Sabrina M Heman-Ackah, MD, DPhil, Oxon, MSE1; Kaasinath Balagurunath1; Selena E Briggs, MD, PhD, MBA, FACS, FACHE2; Rachel Blue, MD3; Christina Jackson, MD3; John Lee, MD3; Georgios Zenonos, MD4; Paul Gardner, MD4; Rania A Mekary, PhD, MS, MSc5; Eduardo Corrales1; Timothy R Smith, MD, PhD, MPH1; 1Mass General Brigham; 2Georgetown MedStar Washington Hospital Center; 3University of Pennsylvania; 4University of Pittsburggh; 5Massachusetts College of Pharmacy and Health Sciences
Background: Facial nerve preservation remains a primary goal in vestibular schwannoma (VS) microsurgery, as postoperative House–Brackmann (HB) scores ≥III significantly impair quality of life. Leveraging an institutional database of 2,500 patients and recent insights into predictors of facial nerve injury, we applied a refined, five-factor, institution-specific machine learning (ML) algorithm to predict HB ≥III outcomes following VS resection. This study builds on our prior multi-institutional ML work, with a focus on clinically relevant variables such as age, obesity, diabetes mellitus (DM), and systemic comorbidities.
Objective: To develop an institution-specific ML algorithm that identifies potentially modifiable predictors of facial nerve injury following VS microsurgery.
Methods: We retrospectively analyzed 2,500 consecutive patients who underwent VS resection via retrosigmoid, translabyrinthine, or middle fossa approaches between 1994–2015. Forty-four demographic, clinical, and radiographic variables were collected. Facial nerve injury was defined as HB ≥III. Based on institutional experience and event distribution, we refined the final model to five predictors, with an 80/20 train–test split (52.2 events in the test set). Random forest, support vector machine (SVM), and RBF-SVM models were trialed; random forest demonstrated superior performance.
Results: Age remained a strong predictor of postoperative injury, consistent with prior multi-institutional findings. In contrast, tumor size emerged as a key predictor in this single-institution analysis, consistent with prior literature. Obesity and DM contributed modestly to model performance. The final random forest model demonstrated strong predictive ability (PPV 80%, NPV 74%, ROC-AUC 0.83, accuracy 76.4%, F1-score 77%).
Conclusion: A machine learning model incorporating age, tumor size, obesity, DM, and systemic comorbidities effectively predicted facial nerve injury after VS resection. This institution-specific approach underscores the value of localized algorithm training to enhance clinical relevance. Importantly, the confirmation of obesity as a predictor highlights a potentially modifiable risk factor, supporting further study into preoperative weight optimization as a strategy to improve facial nerve outcomes.