Establishment and validation of a predictive machine learning model of postoperative long-term diabetes insipidus following transsphenoidal surgery of sellar lesions.
Abstract
Background: Transsphenoidal surgery is the standard treatment for sellar lesions, including pituitary adenomas and related pathologies. Objective: To establish and validate a predictive machine learning model for postoperative long-term diabetes insipidus following transsphenoidal surgery for sellar lesions. Methods: This retrospective analytical study was conducted at Hayatabad medical complex hospital Peshawar December from 2024 to December 2025, included 152 patients who underwent transsphenoidal surgery. Clinical, radiological, and intraoperative variables were collected, including tumor size, suprasellar extension, posterior pituitary bright spot status, and intraoperative cerebrospinal fluid (CSF) leak. Results: The mean age of patients was 45.3 ± 12.6 years, with 57.9% males. Long-term DI developed in 29 (19.1%) patients. Significant predictors included larger tumor size (3.4 ± 0.9 cm vs 2.6 ± 0.8 cm, p<0.001), suprasellar extension (75.9% vs 46.3%, p=0.006), intraoperative CSF leak (62.1% vs 28.5%, p=0.002), and absence of posterior pituitary bright spot (69.0% vs 33.3%, p=0.001). Among the models, the random forest algorithm demonstrated the highest predictive performance with an accuracy of 86.2%, sensitivity of 82.8%, specificity of 88.5%, and AUC of 0.91. Conclusion: Postoperative long-term diabetes insipidus is a relatively common complication following transsphenoidal surgery. Machine learning models, particularly random forest, provide high predictive accuracy and can facilitate early identification of high-risk patients, enabling targeted monitoring and improved postoperative management.