Machine Learning-based Prediction of Long-term Outcomes Following Pituitary Surgery: Analysis of 1,072 Cases at Mayo Clinic

Introduction: While surgical outcomes for pituitary adenomas vary among institutions, predicting individual patient outcomes remains challenging. We investigated the predictors of long-term outcomes following pituitary surgery using both conventional statistics and machine learning (ML) approaches.

Methods: We analyzed 1,072 patients from the Mayo Adenoma of the Pituitary Enterprise Registry (MAPER) database (IRB #21-010809) who underwent pituitary surgery between 2013 and 2023 at three Mayo Clinic sites. The cohort included patients who provided research consent with a median follow-up of 56 months. Primary endpoints were intervention-free rates (IFR) and overall survival (OS). Five different ML models, including Cox proportional hazards, random survival forest, gradient boosting survival, extra survival trees, and survival support vector machine models, were evaluated using C-index and Integrated Brier Score (IBS) as performance metrics.

Results: Patient demographics (median age 52 years, 52.7% female) and surgical details were summarized. Surgical outcomes were stratified by tumor subtypes, demonstrating gross total resection in 65.3% with low complication rates (CSF leak 2.4%, infection 1.0%, visual deterioration 0.7%). IFR and OS were 72% and 98% at 5 years, respectively. For outcome prediction, gradient boosting survival showed optimal performance for IFR (C-index: 0.6235, IBS: 0.1385) while random survival forest performed best for OS (C-index: 0.9447, IBS: 0.1000). ML revealed distinct predictive patterns for better IFR and OS: lower body mass index, smaller tumor size and higher extent of resection were crucial for both outcomes. Additionally, better IFR was primarily influenced by older patient age, endoscopic approach, smaller tumor size, nonfunctioning tumors and no cavernous sinus invasion, while better OS was strongly predicted by younger age, absence of Cushing's disease and female sex.

Conclusion : This comprehensive analysis demonstrates that while tumor subtype influences outcomes, patient-specific factors and surgical variables have distinct impacts on IFR and OS. ML models revealed complex interactions between clinical variables, suggesting that outcome prediction requires integrated assessment of multiple factors. These findings may help guide individualized surgical approach selection and post-operative surveillance strategies.