Machine Learning Approach for Prediction of Outcomes for Surgically Treated Prolactinomas

Introduction: Advanced machine learning (ML) models often outperform logistic regression (LR) in predicting clinical outcomes, yet ML is underutilized for predicting outcomes following transsphenoidal resection of prolactinomas.

Methods: We compared eight ML methods (LR, random forest, support vector, XGBoost, lightGBM, CatBoost, voting, and stacking classifiers) for their ability to predict gross total resection (GTR), short-term remission (STR), long-term remission (LTR), and resolution of preoperative hypopituitarism (RPH) in pathology-confirmed prolactinomas treated at our institution from 2010-2020 (N=50, 64% female, mean age 35.8±13.1 years, 66% macroadenoma). The voting and stacking models used support vector classifier, random forest classifier, and XGBoost classifier as base learners.

Results: For LTR prediction, ML models including lightGBM and random forest reached accuracies of 0.7 (AUC=0.5) and 0.6 (AUC=0.5), respectively. The LR model, in contrast, achieved an accuracy of 0.3. Key predictors for LTR included GTR, postoperative prolactin level, mean tumor axis length, dopamine agonist resistance, and age. For STR prediction, ML models such as XGBoost and voting classifiers (accuracy=0.7, AUC=0.7) slightly outperformed LR (accuracy=0.6, AUC=0.6), with important predictors being GTR, preoperative prolactin level, mean tumor axis length, and age. All models performed well in predicting GTR, with several models having accuracies of 0.9. These models were influenced by factors like cavernous sinus invasion, Knosp grade, Ki-67 index, preoperative prolactin level, mean tumor axis length, and age. For RPH prediction, ML models such as support vector, XGBoost, and CatBoost classifiers (accuracy=0.8, AUC=0.8) slightly outperformed LR (accuracy=0.7, AUC=0.7), with predictors including age, suprasellar extension, mean tumor axis length, and preoperative prolactin level.

Conclusion : These findings suggest that advanced ML models can offer improvements in predictive accuracy over traditional LR likely due to their ability to capture nonlinear relationships between features. These models highlight the potential role of ML in enhancing prognostic precision and supporting individualized treatment strategies in prolactinoma management.