480 Deep Learning Analysis of Early MRI Scans After Severe Traumatic Brain Injury Predicts Long-term Outcomes. (1st April 2022)

• Record Type: Journal Article
• Authors: Pease, Matthew; Arefan, Dooman; Puccio, Ava; Roy, Souvik; Nwachuku, Enyinna; Wu, Shandong; Okonkwo, David O.
• Is Part Of: Neurosurgery. Volume 68(2022)Supplement 1
• Page Start: 118
• Page End: 119
• Publication Date: 2022-04-01
• Subjects: Nervous system -- Surgery -- Periodicals; 617.48005
• DOI: 10.1227/NEU.0000000000001880_480
• Languages: English
• ISSNs: 0148-396X

Abstract

Abstract : INTRODUCTION: Nearly three quarters of deaths in clinical trials of severe traumatic brain injury (TBI) occur due to withdrawal of life sustaining therapies, often due to perceived poor neurological prognosis. Most of these deaths occur within the first week when physicians historically lack enough accuracy to make these decisions. Prognostic models, such as the International Mission for Prognosis and Analysis of Clinical Trials in TBI, are designed to prognose outcomes upon admission and fail to utilize clinical information accumulated after admission. Quantitative prognostic models capturing this informaiton are needed to help guide care decisions. METHODS: We built our model using 266 patients from a prospective database of severe TBI patients from 2002 through 2018 with an MRI obtained 1 to 9 days post-trauma. We developed a customized, multi-channel, deep convolutional neural network model with an AlexNet backbone and pre-trained with ImageNet. The cohort was split into 80% training, 10% validation, and 10% independent testing set. We built separate models using T2 FLAIR and DWI sequences. To reduce the heterogeneity of imaging after a decompressive hemicraniectomy (DHC), we only used DHC patients for model training. RESULTS: No adverse clinical events occurred while obtaining early MRIs. The overall mortality was 25% (68) and 20% (34) in the 170 patients who did not have a DHC. In the independent test cohort, the T2 FLAIR model performed best, with an area under the curve (AUC) of 0.80 (95% Confidence Interval: 0.57-0.94) and accuracy of 83%. The DWI model had an AUC of 0.64 (0.50 – 0.79) with an accuracy of 78%. With a specificity of 100% (i.e., never recommending withdrawal of care in a patient who would otherwise survive), the sensitivity for mortality was 57%. For identifying patients with favorable recovery potential, the sensitivity was 83% with a specificity of 80%. CONCLUSION: Early MRIs are safe to obtain in the severe TBI patient population. Deep learning analysis of these MRIs can predict 6-month mortality to help guide neurological prognostication.