Artificial intelligence-based prediction of cerebrospinal fluid leak in skull base surgery: a scoping review

Cerebrospinal fluid (CSF) leak remains one of the most significant complications of skull base surgery, particularly in endoscopic endonasal and transsphenoidal approaches. Accurate risk stratification may improve surgical planning and reconstruction strategies. Artificial intelligence (AI)-based models have been increasingly investigated for CSF leak prediction; however, the evidence remains methodologically heterogeneous. This scoping review aimed to map the current literature and summarize data sources, modeling approaches, validation strategies, and reported performance.

This scoping review followed PRISMA-ScR guidance. Eligibility was structured using the Population-Concept-Context framework. PubMed/MEDLINE, Scopus, and Web of Science were searched for English-language studies published from 2019 to 2025. Two authors independently screened titles/abstracts and full texts and independently charted data using a standardized table; disagreements were resolved by consensus. Eligible original human studies applied AI/ML to predict perioperative surgical CSF leak or to derive an imaging biomarker for identifying spontaneous lateral skull-base CSF leak/encephalocele relevant to surgical assessment. No formal protocol was prepared or registered.

Fifty-five records were identified, and nine primary studies met the eligibility criteria after duplicate removal and full-text assessment. Six studies predicted intraoperative CSF leak, two predicted postoperative leak, and one evaluated a deep-learning-derived CT biomarker for identifying spontaneous lateral skull-base CSF leak or encephalocele relevant to surgical assessment. Three studies used classical machine-learning methods, four used predominantly deep neural-network or convolutional-neural-network approaches, one compared classical models with a feedforward neural network, and one used a BERT-based natural-language-processing model. One study reported prospective temporal validation; none reported independent external validation at another institution. Among studies reporting AUC, values ranged from 0.69 to 1.00; accuracy, where reported, ranged from 70% to 98.33%.

AI-based prediction of CSF leak is a promising but evolving field. Methodological heterogeneity and the absence of independent external validation constrain clinical translation. Future studies should prioritize standardized outcomes, multicenter datasets, and prospective validation to support integration into clinical decision-making.