Deep Learning to Predict TBI Outcomes in the Low Resource Setting. (16th November 2020)

• Record Type: Journal Article
• Authors: Adil, Syed M; Elahi, Cyrus; 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_488
• 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, accurate patient prognostication is both difficult and essential. Deep learning-a branch of machine learning using neural networks with multiple hidden layers-has the potential to predict outcomes better than other machine learning algorithms and capture complex non-linear patterns. METHODS: Data from TBI patients of all ages were prospectively collected at Mulago National Referral Hospital in Kampala, Uganda, from 2016 to 2020. We designed the 3 aforementioned models to predict good versus poor outcome at hospital discharge. The DNN included four hidden layers. Predictors included 13 clinical variables easily acquired on admission-spanning demographics, physical exam, presence of polytrauma or seizures, and mechanism of injury-and whether or not the patient received surgery. Model performance was assessed using 5-fold cross-validation. We calibrated the model using Platt scaling. RESULTS: Ultimately, 2164 patients were included for model training and a subset of 1677 for model testing, of which 12% had poor outcomes. The mean age was 28 -± 15 years and 85% were male. The mean admission Glasgow Coma Score (GCS) was 12.4 ± 2.9. Twenty-eight percent of patients received surgical intervention. The DNN demonstrated the highest area under the receiver operating characteristic curve (AUC) at 93.2% with an F1 score of 0.70, followed by the SNN at 92.7% and 0.68, and finally the LRnet at 92.1% and 0.64, respectively. CONCLUSION: We present one of the first uses of deep learning to predict outcomes after TBI in the LMIC setting. The model slightly outperforms both SNN and LRnet on composite metrics. All models performed well. Before implementation, the model should be externally validated on other LMIC data. Future studies should continue optimization of DNN model architecture and illuminate the individual treatment effect associated with surgery in these predictions, with the ultimate aim of enhancing surgical decision making in the low-resource setting.