Machine Learning-Based Predictive Analytics in Transsphenoidal Surgery for Pituitary Adenoma

Background: Transsphenoidal surgery has become the standard of care for most pituitary adenomas, and is a relatively safe procedure with a high success rate. Personalized medicine has moved to the forefront of medical research in the past decade, and can potentially impact clinical practice. For example, preoperative identification of patients at high risk for subtotal resectability or intraoperative cerebrospinal fluid (CSF) leaks may allow surgeons to better inform patients on the likelihood of outcomes and adverse events, adjust surgical targets, and create potential for prevention of complications. Recently, machine learning algorithms for predictive analytics have demonstrated superiority to conventional statistical modeling in a range of fields.

Objective: To develop robust predictive analytics that assist surgeons in preoperatively identifying the likelihood of a particular patient to experience gross-total resection (GTR), intraoperative CSF leaks, overall complications, improvement or worsening of preoperative hormonal deficits, and biochemical remission in secreting adenomas.

Methods: From a prospective pituitary adenoma registry, patients were identified. For each of the five outcomes, a machine learning-based prediction model was trained and internally validated. Data were randomly split into training, validation, and testing set in a 70:15:15 ratio. Class imbalance in the training set was handled by applying synthetic minority oversampling (SMOTE). For each outcome, we trained models based on deep neural networks, random forests, and extreme gradient boosting, and selected the best model based on area under the curve (AUC) on the validation set. The final models were then evaluated on the test set. Various countermeasures were taken to minimize overfitting. Missing data were imputed using single imputation. We also aimed to identify traditional, independent predictors for the outcomes using univariate testing and logistic regression modeling.

Results: Data from 154 patients were available. Among all five models, we observed high performance measures at internal validation, indicating that the machine learning-based prediction models were robust, and outperformed conventional predictive analytics such as logistic regression. Specifically, AUC of up to 0.96, accuracy of up to 91%, sensitivity and specificity of up to 94%/89%, positive predictive value (PPV) and negative predictive value (NPV) of up to 89%/94%, and F1 score of up to 0.91. Deep neural networks performed best. Notably, for all outcomes, no reliable independent predictors were identified using conventional statistical methods.

Conclusion: Machine learning helps identify patients prone to certain outcomes, possibly enabling better, more personalized treatment. External validation and the development of a freely accessible web-based tool integrating all five models will demonstrate the clinical practicality of such predictive analytics in patients undergoing surgery for pituitary adenoma.

No conflict of interest has been declared by the author(s).