A Novel BERT-Based Machine Learning Approach for Enhanced CSF Leak Prediction in Endoscopic Endonasal Skull Base Surgery.

Objectives: To evaluate the performance of the BERT (bidirectional encoder representations from transformers) model in predicting cerebrospinal fluid (CSF) leaks and compare it with traditional logistic regression analysis. Methods: In this study, we employed a machine learning-based natural language processing (NLP) model, specifically BERT, and compared its performance to conventional statistical logistic regression in predicting CSF leaks. We analyzed all cases of skull base pathologies treated by a multidisciplinary team specializing in rhinology and skull base surgery, and neurosurgery at a single center between March 2015 and July 2020. The dataset included the following factors: (1) demographics, (2) perioperative clinical CSF leak indicators, (3) pathology-related factors, (4) surgical factors, and (5) perioperative CT scan features. Results: The BERT model outperformed the traditional logistic regression model in predicting CSF leaks, achieving an AUC of 1.0000, sensitivity of 1.0000, specificity of 0.9808, positive predictive value (PPV) of 0.8889, negative predictive value (NPV) of 1.0000, and an F1 score of 0.9657. In contrast, the logistic regression model yielded an AUC of 0.847, with a sensitivity of 0.2143, specificity of 0.9060, PPV of 0.4471, and NPV of 0.7649. Conclusion: BERT NLP model outperforms traditional logistic regression in predicting cerebrospinal fluid leaks after endoscopic skull base surgery, demonstrating superior accuracy with qualitative clinical data, enhancing risk stratification and decision-making.