Outcome Prediction in Patients with Severe Traumatic Brain Injury Using Deep Learning from Head CT Scans

Abstract

Background After severe traumatic brain injury (sTBI), physicians use long-term prognostication to guide acute clinical care yet struggle to predict outcomes in comatose patients. Purpose To develop and evaluate a prognostic model combining deep learning of head CT scans and clinical information to predict long-term outcomes after sTBI. Materials and Methods This was a retrospective analysis of two prospectively collected databases. The model-building set included 537 patients (mean age, 40 years ± 17 [SD]; 422 men) from one institution from November 2002 to December 2018. Transfer learning and curriculum learning were applied to a convolutional neural network using admission head CT to predict mortality and unfavorable outcomes (Glasgow Outcomes Scale scores 1-3) at 6 months. This was combined with clinical input for a holistic fusion model. The models were evaluated using an independent internal test set and an external cohort of 220 patients with sTBI (mean age, 39 years ± 17; 166 men) from 18 institutions in the Transforming Research and Clinical Knowledge in Traumatic Brain Injury (TRACK-TBI) study from February 2014 to April 2018. The models were compared with the International Mission on Prognosis and Analysis of Clinical Trials in TBI (IMPACT) model and the predictions of three neurosurgeons. Area under the receiver operating characteristic curve (AUC) was used as the main model performance metric. Results The fusion model had higher AUCs than did the IMPACT model in the prediction of mortality (AUC, 0.92 [95% CI: 0.86, 0.97] vs 0.80 [95% CI: 0.71, 0.88]; P < .001) and unfavorable outcomes (AUC, 0.88 [95% CI: 0.82, 0.94] vs 0.82 [95% CI: 0.75, 0.90]; P = .04) on the internal data set. For external TRACK-TBI testing, there was no evidence of a significant difference in the performance of any models compared with the IMPACT model (AUC, 0.83; 95% CI: 0.77, 0.90) in the prediction of mortality. The Imaging model (AUC, 0.73; 95% CI: 0.66-0.81; P = .02) and the fusion model (AUC, 0.68; 95% CI: 0.60, 0.76; P = .02) underperformed as compared with the IMPACT model (AUC, 0.83; 95% CI: 0.77, 0.89) in the prediction of unfavorable outcomes. The fusion model outperformed the predictions of the neurosurgeons. Conclusion A deep learning model of head CT and clinical information can be used to predict 6-month outcomes after severe traumatic brain injury. © RSNA, 2022 Online supplemental material is available for this article. See also the editorial by Haller in this issue.

Graphical abstract

Figure 1:

Consolidated Standards of Reporting Trials diagram for (A) University of Pittsburgh Medical Center and (B) Transforming Research and Clinical Knowledge in Traumatic Brain Injury cohorts. For patients who were missing 6-month outcomes, the 3- or 12-month outcome was substituted in place of the 6-month outcome, if available for model prediction. CTH = CT of the head, GCS = Glasgow Coma Scale, GOS = Glasgow Outcome Scale, TBI = traumatic brain injury.

Figure 2:

Outline of deep learning modeling to predict long-term outcomes in patients with severe traumatic brain injury based on radiographic and clinical information available in the emergency department. (A) An analysis of multimodal data, including a customized convolutional neural network (CNN) structure for modeling CT imaging data (imaging model) and a clinical model, was performed to generate a holistic prediction of the long-term outcomes (fusion model). (B) The customized CNN imaging model was structured using AlexNet backbone. The size of kernels in the input layer was changed from 11 × 11 × 3 in AlexNet to 11 × 11 × 7 in the customized model. Transfer learning was applied for all learnable layers, except for the last fully connected layer (FC8). For the input layer with seven channels, each of the available 96 kernels in the input layer of a pretrained AlexNet were averaged over the third dimension (three red-green-blue channels), and then the weights to each of the seven channels of the kernels in the input layer of the CNN model were transferred. Conv = convolutional layer, FC = fully connected layer..

Figure 3:

Comparison of performance of imaging, fusion, and International Mission on Prognosis and Analysis of Clinal Trials in Traumatic Brain Injury (IMPACT)-fusion models with IMPACT for survival and unfavorable outcomes. Receiver operating characteristic curves compare (A) mortality and (B) unfavorable outcomes for the University of Pittsburgh Medical Center data set and (C) mortality and (D) unfavorable outcomes for the Transforming Research and Clinical Knowledge in Traumatic Brain Injury (TRACK-TBI) validation. Sensitivity and specificity for attending neurosurgeon predictions are reported in A and B, both as an average and for the individual attending neurosurgeon with 1 (A1), 5 (A5), and 25 (A25) years of experience.

Figure 4:

Example predictions by fusion model on University of Pittsburgh Medical Center patients. (A) Correct prediction in a 44-year-old man who was involved in an unrestrained motor vehicle collision. He underwent emergent decompressive hemicraniectomy (DHC), had bilateral lung injuries, and ultimately developed a pulmonary embolism with difficulty oxygenating on posttrauma day 6. His care was withdrawn, and he died. The model correctly predicted mortality. (B) Incorrect prediction in a 57-year-old woman who was in a motor vehicle collision and underwent DHC. The model predicted she would die, but she had a Glasgow Outcomes Scale of 3 at 2 years after trauma. She lived in a nursing home and was dependent on others for most daily living activities. (C) Incorrect prediction in a 28-year-old man who was in a motorcycle collision and had a minor head injury with intraventricular hemorrhage. Several weeks after trauma, he developed Klebsiella ventriculitis and pneumonia that led to an episode of severe hypotension. He subsequently developed malignant cerebral edema and died by brain death criteria. While the model predicted this patient would survive, this scenario highlights the difficulty of predicting outcomes based on information available in the emergency department, as events later in the patient’s course affect outcomes.