Machine learning algorithms for predicting delayed hyponatremia after transsphenoidal surgery for patients with pituitary adenoma

This study aimed to develop and validate machine learning (ML) models to predict the occurrence of delayed hyponatremia after transsphenoidal surgery for pituitary adenoma. We retrospectively collected clinical data on patients with pituitary adenomas treated with transsphenoidal surgery between January 2010 and December 2020. From January 2021 to December 2022, patients with pituitary adenomas were prospectively enrolled. We trained seven ML models to predict delayed hyponatremia using the clinical variables in the training set. The final model was internally validated using a test set and a prospective dataset. The SHapley Additive exPlanations (SHAP) algorithm was used to determine the significance of each variable in the occurrence of delayed hyponatremia. In the training dataset, the best predictive performance was observed for XGBoost (area under the ROC curve; AUC = 0.821), followed by Random Forest (AUC = 0.8), Logistic Regression (AUC = 0.793), Support Vector Machine (AUC = 0.776), naïve Bayes (AUC = 0.774), K-Nearest Neighbors (AUC = 0.742), and Decision Tree (AUC = 0.717). The AUC of the XGBoost model for the test and prospective datasets are 0.831 and 0.785, respectively. The differences in pituitary stalk deviation angle, the "measurable pituitary stalk" length before and after surgery, and blood sodium concentration between preoperative and postoperative day 2 were important variables for predicting delayed hyponatremia as determined by the SHAP algorithm. The XGBoost model was best able to predict delayed hyponatremia after transsphenoidal surgery for pituitary adenomas. The differences in pituitary stalk deviation angle, pre- versus postoperative "measurable pituitary stalk" length, and pre- versus postoperative day 2 blood sodium concentrations were important variables for predicting delayed hyponatremia.