A supervised machine-learning algorithm predicts intraoperative CSF leak in endoscopic transsphenoidal surgery for...

Area of Science:

• Neurosurgery
• Machine Learning
• Medical Informatics

Background:

• Cerebrospinal fluid (CSF) leakage is a critical complication following endoscopic transnasal transsphenoidal surgery (E-TNS) for pituitary adenomas (PAs).
• Predicting and mitigating CSF leakage risk is crucial for improving patient outcomes and reducing mortality.

Purpose of the Study:

• To develop and validate a supervised machine learning (ML) model for predicting intraoperative CSF leakage during E-TNS for PAs.
• To compare the performance of various ML algorithms in identifying patients at high risk for CSF leakage.

Main Methods:

• A retrospective cohort of 238 patients undergoing E-TNS for PAs was analyzed.
• Several ML models were trained, with the top five evaluated on a hold-out test set.
• The best-performing classifier was prospectively validated on a separate cohort of 35 patients.

Main Results:

• The random forest (RF) classifier demonstrated superior performance, achieving an AUC of 0.84 with high sensitivity (86%) and specificity (88%).
• Key predictors identified included non-secreting status, older age, tumor dimensions, invasiveness, volume, ICD, and R-ratio.
• Prospective validation confirmed robust performance with an AUC of 0.81 and high negative predictive value (95%).

Conclusions:

• The RF classifier is a highly effective tool for predicting intraoperative CSF leakage in patients undergoing E-TNS for PAs.
• ML models, particularly RF, offer advantages over traditional statistical methods for predicting surgical outcomes in complex patient populations.
• Implementing RF models can enhance patient management, reduce preventable morbidity, and lower healthcare costs.