Predictive modeling of postoperative hyponatremia after pituitary adenoma surgery

Guoge Li; Xiaotong Li; Yaowei Ding; Jialu Sun; Yuxin Chen; Siqi Wang; Zhijun Yang; Guojun Zhang

doi:10.1016/j.clineuro.2025.109020

Predictive modeling of postoperative hyponatremia after pituitary adenoma surgery

Clin Neurol Neurosurg. 2025 Jun 21:256:109020. doi: 10.1016/j.clineuro.2025.109020. Online ahead of print.

Authors

Guoge Li¹, Xiaotong Li¹, Yaowei Ding¹, Jialu Sun¹, Yuxin Chen¹, Siqi Wang¹, Zhijun Yang², Guojun Zhang³

Affiliations

¹ Department of Clinical Diagnosis, Laboratory of Beijing Tiantan Hospital, Capital Medical University, Beijing, China; NMPA Key Laboratory for Quality Control of In Vitro Diagnostics, Beijing, China; Beijing Engineering Research Center of Immunological Reagents Clinical Research, Beijing, China.
² Department of Neurosurgery of Beijing Tiantan Hospital, Capital Medical University, Beijing, China. Electronic address: zhijunyang@ccmu.edu.cn.
³ Department of Clinical Diagnosis, Laboratory of Beijing Tiantan Hospital, Capital Medical University, Beijing, China; NMPA Key Laboratory for Quality Control of In Vitro Diagnostics, Beijing, China; Beijing Engineering Research Center of Immunological Reagents Clinical Research, Beijing, China. Electronic address: guojun.zhang@ccmu.edu.cn.

PMID: 40578223
DOI: 10.1016/j.clineuro.2025.109020

Abstract

Objective: To improve the prediction of postoperative hyponatremia after pituitary surgery by comparing six machine learning (ML) models.

Methods: We analyzed 777 patients post-pituitary surgery, with 118 developing hyponatremia, and validated our findings in an external cohort of 183 patients (28 hyponatremia cases). Six ML models were developed using pre- and postoperative clinical and laboratory data, with predictive accuracy assessed via the area under the ROC curve (AUC). Interpretability techniques included permutation feature importance (PFI), local interpretable model-agnostic explanations (LIME), and Shapley Additive exPlanations (SHAP).

Results: Thirteen key features were identified for ML model construction. The Random Forest (RF) model showed the highest predictive accuracy with AUCs of 0.88 in internal validation datasets. PFI analysis identified desmopressin dosage, postoperative day 1 and day 2 serum sodium levels, and surgical duration as the most significant influential variables. LIME and SHAP corroborated PFI's local feature importance findings.

Conclusions: The RF-based ML model offers accurate predictions of postoperative hyponatremia by integrating personalized patient data, potentially enhancing patient outcomes through rapid hyponatremia prediction.

Keywords: Machine Learning; Pituitary Adenoma; Postoperative hyponatremia; Prediction.