Automated Prediction of Intraoperative Pneumocephalus During Deep Brain Stimulation Surgery
Taylor Edwards1, Wasif Khan2, Ruogu Fang2, Justin Hilliard3, Joshua Wong4
1University of Florida College of Medicine, 2University of Florida, 3University of Florida College of Medicine - Neurosurgery, 4University of Florida College of Medicine - Neurology
Objective:
To develop a risk stratification algorithm for intraoperative pneumocephalus using artificial intelligence (AI).
Background:
Deep brain stimulation (DBS) is a safe, effective therapy for conditions including Parkinson’s Disease, Essential tremor, and dystonia. Placement of DBS leads can be affected by intraoperative brain shift caused by air entering the surgical cavity or leakage of cerebrospinal fluid, resulting in a condition called pneumocephalus. When pneumocephalus occurs, brain structures move from locations identified in preoperative imaging, possibly resulting in suboptimal lead placement.
Design/Methods:
We conducted a retrospective review of DBS surgical patients between 1/2022 and 2/2025, collecting patient demographics, clinical background, DBS hardware information, and neuroimaging. We measured pneumocephalus from intraoperative CT immediately after DBS lead placement using a semi-automated segmentation technique via 3D Slicer. We applied voxel-based morphometry to the preoperative MRI brain using Freesurfer to generate an anatomical list of features for AI analyses, and trained binary classification models to predict the risk of developing significant pneumocephalus during DBS surgery. Our model used a standardized 80-20 training-testing split and model performance was evaluated using a 5-fold cross-validation. Our primary performance metric was the area under the receiver operating curve (AUROC). Secondary metrics were the F1 score, precision, recall and balanced accuracy. 145 patients were included. We defined significant pneumocephalus as any volume greater than 3.000 cm3.
Results:
60 patients had significant pneumocephalus while 85 patients did not. The mean (SD) pneumocephalus volume was 4.54214 cm3 (6.09216 cm3). Our best AI model was able to predict the development of significant pneumocephalus with an AUROC of 0.72, a balanced accuracy of 72%, and F1 score of 0.71. There was no data leakage.
Conclusions:
We demonstrate a machine learning model that can reasonably predict the development of significant pneumocephalus during DBS surgery. This tool has the potential to facilitate intraoperative decision making for more consistent patient outcomes.
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.