Predicting the Impact of Neurosurgery on TBI Patients in the Low Resource Setting: A Machine Learning Approach in Uganda. (16th November 2020)

• Record Type: Journal Article
• Authors: Adil, Syed M; Elahi, Cyrus; Gramer, Robert; Spears, Charis A; Fuller, Anthony; Haglund, Michael M; Dunn, Timothy
• Is Part Of: Neurosurgery. Volume 67(2010)Supplement 1
• Publication Date: 2020-11-16
• Subjects: Nervous system -- Surgery -- Periodicals; 617.48005
• DOI: 10.1093/neuros/nyaa447_467
• Languages: English
• ISSNs: 0148-396X

Abstract

Abstract: INTRODUCTION: Traumatic brain injury (TBI) disproportionately affects low- and middle-income countries (LMICs). In these low-resource settings, effective triage of TBI patients-including the decision of whether or not to perform neurosurgery-is critical in optimizing both patient outcomes and healthcare resource utilization. METHODS: Data from TBI patients of all ages were prospectively collected at Mulago National Referral Hospital in Kampala, Uganda, from 2016 to 2019. Seven different machine learning models (based on 1 linear and 6 non-linear algorithms) designed to predict good vs poor outcome near hospital discharge were developed and internally validated using 5-fold cross-validation. Predictors included clinical variables easily acquired on admission-demographics, physical exam, and mechanism of injury-and whether or not the patient received surgery. Using the elastic-net regularized logistic regression model (GLMnet), the probability of poor outcome was calculated for each patient both with and without surgery (quantifying the "treatment benefit"). A relative treatment benefit was then calculated, equaling this benefit of surgery divided by the probability of bad outcome with no surgery. Predictions were calibrated using Platt scaling. RESULTS: Ultimately, 1766 patients were included. Areas under the receiver operating characteristic curve (AUCs) ranged from 81.7% (k-nearest neighbors) to 88.0% (random forest). The GLMnet had the second-best AUC at 87.7%. For the entire cohort, the median relative treatment benefit was 37.6% (IQR, 31.0% to 46.0%); similarly, in just those receiving surgery, it was 38.0% (IQR, 31.4% to 47.0%). The top four variables promoting good outcomes in the GLMnet model were high GCS, being fully alert, having both pupils reactive, and receiving surgery. CONCLUSION: We provide the first deployable machine learning-based model to predict TBI outcomes with and without surgery in LMICs, thus enabling more effective surgical decision making in the resource-limited setting. Currently, patients are not being optimally chosen for neurosurgical intervention. Future studies should externally validate the model, improve model performance by combining data across countries, and explore use of more advanced algorithms.