Deep Learning to Predict Outcome in Severe Traumatic Brain Injury

HomeRadiologyVol. 304, No. 2 PreviousNext Reviews and CommentaryFree AccessEditorialDeep Learning to Predict Outcome in Severe Traumatic Brain InjurySven Haller Sven Haller Author AffiliationsFrom the Centre d'Imagerie Médicale de Cornavin, Place de Cornavin 18, 1201 Geneva, Switzerland; Department of Surgical Sciences, Radiology, Uppsala University, Uppsala, Sweden; Faculty of Medicine of the University of Geneva, Geneva, Switzerland; Department of Radiology, Beijing Tiantan Hospital, Capital Medical University, Beijing, China.Address correspondence to the author (email: [email protected]).Sven Haller Published Online:Apr 26 2022https://doi.org/10.1148/radiol.220412MoreSectionsPDF ToolsImage ViewerAdd to favoritesCiteTrack CitationsPermissionsReprints ShareShare onFacebookTwitterLinked In See also the article by Pease and Arefan et al in this issue.Sven Haller, MD, is a neuroradiologist and medical director at Centre d'Imagerie Médicale de Cornavin in Geneva, Switzerland, and visiting professor at the University of Uppsala, Sweden, and Tiantan Hospital in Beijing, China. He has special interest in advanced neuroimaging techniques, including functional MRI, real-time functional MRI neurofeedback, arterial spin labelling, and susceptibility-weighted imaging, in neurodegenerative and neurovascular diseases. He has received multiple international scientific distinctions and has leadership roles in the European Society of Neuroradiology.Download as PowerPointOpen in Image Viewer Severe traumatic brain injury (sTBI) is a devastating event for patients and relatives. Outcome prediction in patients with sTBI is not a novel concept; however, the available approaches are not ready for clinical application. Instead, approaches are often variable between centers and are operator dependent.In this issue of Radiology, Pease and Arefan et al describe a deep learning (DL) approach combining basic clinical parameters and admission CT scans to predict 6-month outcomes in patients with sTBI (1).This retrospective analysis of two prospectively collected databases first trained several different models in 537 patients from one institution. The DL model performance was evaluated in an independent internal test set and an additional external test set of 220 patients from 18 institutions in the Transforming Research and Clinical Knowledge in Traumatic Brain Injury (TRACK-TBI) study. This DL-based outcome prediction was compared with the International Mission on Prognosis and Analysis of Clinical trials in Traumatic Brain Injury (IMPACT) clinical outcome prediction score and the assessment of attending neurosurgeons. Of note, although IMPACT attempts to predict outcomes for patients by using emergency department information, this score was designed primarily for clinical trials rather than clinical applications (2,3).The best-performing DL model included initial clinical and head CT information. When considering only the single-institution internal data set, this combined DL model successfully predicted mortality (area under the receiver operating characteristic curve [AUC], 0.92) and unfavorable outcomes (AUC, 0.88) at 6 months. When considering the external validation set, performance of the model decreased to an AUC of 0.80 concerning mortality, which showed no evidence of a difference compared with IMPACT (AUC, 0.83; P = .05) and was higher than that of the attending neurosurgeons.Overall, the presented DL-based approach (1) is a clinically relevant and important step toward clinically useful, standardized, and operator-independent outcome prediction in the sTBI setting, and the authors are to be commended for achieving this difficult task using independent training and testing sets in a clinically relevant implementation.A few points might be considered to further improve the clinically highly relevant outcome prediction in the sTBI setting. First, the fact that model performance decreased when transferred from the training set to the independent test set is a very common problem in machine learning. Model performance might be improved by using a larger and more heterogeneous multicenter data set for model training that better matches the more variable multicenter data set of 18 institutions of the external TRACK-TBI test set. This would also better reflect the variability of CT data acquisition in real clinical settings in different institutions. Second, in this study, the attending neurosurgeons had access to the same clinical information and CT scans as the model, and they made binary predictions for mortality and outcomes at 6 months. This artificial research setting does not fully correspond to the real clinical setting in which the neurosurgeons also see and clinically examine the patients, thus providing additional real-world clinical decision parameters for neurosurgeons. This suggests that neurosurgeons might perform better in a more familiar real-world scenario than in the artificial research setting of the presented study. Third, CT scans were evaluated by neurosurgeons rather than radiologists. This is not to question the clinical experience of neurosurgeons, but it might be that radiologists can provide a more accurate assessment of CT scans.In summary, this study implies that the performance of the automatic classifier might be further enhanced by using a multicenter data set for training. On the other hand, it also implies that attending neurosurgeons (and radiologists) might perform better in a real-world scenario than in this artificial research setting. Altogether, this implies that future prospective studies and additional real-world studies are warranted to compare the proposed DL automatic classifier with neurosurgeons and radiologists in a more realistic clinical setting.Disclosures of conflicts of interest: Consulting fees from Wyss Center For Bio And Neuroengineering and Geneva Spineart; lectures for GE; expert testimony for Varia; European Society of Neuroradiology artificial intelligence group leader.References1. Pease M, Arefan D, Barber J, et al. Outcome prediction in patients with severe traumatic brain injury using deep learning from head CT scans. Radiology 2022;304(2):385–394. Link, Google Scholar2. Steyerberg EW, Mushkudiani N, Perel P, et al. Predicting outcome after traumatic brain injury: development and international validation of prognostic scores based on admission characteristics. PLoS Med 2008;5(8):e165; discussion e165. Crossref, Medline, Google Scholar3. Letsinger J, Rommel C, Hirschi R, Nirula R, Hawryluk GWJ. The aggressiveness of neurotrauma practitioners and the influence of the IMPACT prognostic calculator. PLoS One 2017;12(8):e0183552. Crossref, Medline, Google ScholarArticle HistoryReceived: Feb 21 2022Revision requested: Mar 9 2022Revision received: Mar 10 2022Accepted: Mar 14 2022Published online: Apr 26 2022Published in print: Aug 2022 FiguresReferencesRelatedDetailsAccompanying This ArticleOutcome Prediction in Patients with Severe Traumatic Brain Injury Using Deep Learning from Head CT ScansApr 26 2022RadiologyRecommended Articles Amyloid PET: A Potential Biomarker for Individuals with Mild Traumatic Brain InjuryRadiology2023Volume: 307Issue: 5Arterial Spin Labeling Perfusion of the Brain: Emerging Clinical ApplicationsRadiology2016Volume: 281Issue: 2pp. 337-356Resting-State Functional MRI Changes in Normal Human AgingRadiology2022Volume: 304Issue: 3pp. 633-634Emerging Perspectives on MRI Application in Multiple Sclerosis: Moving from Pathophysiology to Clinical PracticeRadiology2023Volume: 307Issue: 5Outcome Prediction in Patients with Severe Traumatic Brain Injury Using Deep Learning from Head CT ScansRadiology2022Volume: 304Issue: 2pp. 385-394See More RSNA Education Exhibits Deep-Brain: A Cutting-edge Concept for Outstanding Functional Resolution in fMRIDigital Posters2020Imaging for Epilepsy Surgery: How to Assist the Epileptologist and NeurosurgeonDigital Posters2019Toolkit for Functional MRI Assessment of Peritumoral Non-Enhancing Areas in Brain LesionsDigital Posters2019 RSNA Case Collection Non-accidental anoxic brain injury RSNA Case Collection2021Cerebral Arteriovenous MalformationRSNA Case Collection2021High Pressure Injection InjuryRSNA Case Collection2021 Vol. 304, No. 2 Metrics Altmetric Score PDF download