Feasibility study of an interpretable machine based model to predict postoperative delayed hyponatremia in pituitary adenoma patients

Objective To investigate the feasibility of constructing an interpretable machine model for predicting delayed hyponatraemia after transsphenoidal pituitary tumour surgery based on the XGBoost algorithm.Methods A total of 168 pituitary adenoma patients who were admitted to the Affiliated Hospital of Xuzhou Medical University from August 2018 to August 2021, and treated via neuroendoscopic transsphenoidal approach were enrolled and their clinical and laboratory data were collected. Then, the XGboost algorithm was used to establish a predictive model and the factors influencing the occurrence of delayed hyponatremia after transsphenoidal pituitary tumor surgery were interpreted based on SHAP algorithm.Results Among the 168 patients with pituitary adenoma, 35 developed postoperative delayed hyponatraemia and 133 did not develop postoperative delayed hyponatraemia. The established predictive model performed well on all evaluation criteria, with the R² score of 0.94. The final analysis based on SHAP values revealed that preoperative sodium levels, thyroid hormone levels and age were the main features in predicting the occurrence of delayed hyponatraemia after transsphenoidal pituitary tumour surgery. The interpretability of the model was further validated through analysis of two individual cases.Conclusions We established the XGBoost-based machine learning model to predict the likelihood of delayed hyponatraemia after transsphenoidal pituitary tumour surgery, and utilize SHAP algorithm to improve the interpretability of the model. These findings can provide reference and guidance for clinical work, in order to effectively reduce postoperative morbidity and improve patient safety.