2025 Proffered Presentations
S295: MACHINE LEARNING-BASED PREDICTION MODEL OF CSF LEAK IN SKULL BASE RECONSTRUCTION: A NOVEL APPROACH.
Rayan AlFallaj, MBBS; Abdulaziz Alrasheed, MBBS; King Saud University medical city
Background: As artificial intelligence continues to advance in healthcare, the integration of machine learning techniques that can handle vast amounts of data offers the opportunity to develop prediction models By serving as a predictive tool, machine learning can assist surgeons in enhancing planning and preparation for skull base surgery reconstruction.In this work, we propose a novel approach based on language models for CSF leak prediction. Unlike existing approaches our solution exploits the power of language models to represent the different variables related to CSF leak using natural language embedding.
Methodology: we adopt the BERT (Bidirectional Encoder Representations from Transformers) model as a backbone. This model is trained in end-to-end manner via the backpropagation algorithm. In a split of Train/Test (70%/30%) on a dataset composed of 120 patients records with 52 input variables distributed over several factors such as age, gender, surgery location, pathology size, preoperative diversion, size of pathology, preoperative hydrocephalus, BMI categories, type of reconstruction material, and tumor location categories. the model was compared to traditional statstical models: The logistic regression model, trained on the upsampled data, with its predictive accuracy and discriminatory power, validated through performance metrics such as the Receiver Operating Characteristic (ROC) curve and the concordance statistic (kappa statistic)
Results: This novel model yields an accuracy: 0.9722, Precision: 0.9848, Recall: 0.8750, and F1-score: 0.9209 compared to traditional logistic regression model: accuracy: 0.725, and 0.4471 positive predictive value and 0.0561 detection rate.
Conclusion: In this study, we introduce a groundbreaking machine learning-based prediction model that significantly outperforms traditional statistical methods. This model can effectively identify the risk factors associated with cerebrospinal fluid (CSF) leaks in endoscopic skull base surgery.