Current methods for radiological classification in traumatic brain injury (TBI), such as the Marshall and Rotterdam score, provide an incomplete description of intracranial lesion burden, rely on time-consuming manual assessments by experts, and are prone to intra- and interrater disagreement. To circumvent these limitations, we previously proposed the Brain Lesion Analysis and Segmentation Tool for Computed Tomography (BLAST-CT), a deep learning-based method using convolutional neural networks (CNNs) to perform multiclass, voxel-wise segmentation and quantification of TBI lesions on CT in an automated fashion. In this study, we expand on our previous work by (1) optimizing the performance of our model using additional training data from CENTER-TBI, (2) externally validating the findings reported in our internal development study by applying our algorithm to an independent imaging dataset from the Prophylaxis for Venous Thromboembolism in Severe Traumatic Brain Injury (PROTEST) multicenter randomized controlled trial. A total of 680 scans from CENTER-TBI were annotated by neuroradiological experts and used to retrain the CNN, creating BLAST version 2.0. Traumatic lesions were subdivided into four distinct classes: intraparenchymal hemorrhage (IPH), extra-axial hemorrhage (EAH), intraventricular hemorrhage (IVH), and perilesional edema. Fifty-one scans from PROTEST were manually annotated to obtain ground-truth lesion labels on an independent imaging dataset. The same PROTEST scans were then contemporaneously run through versions 1.0 and 2.0 of BLAST-CT to evaluate the performance change resulting from the optimization training procedure while calculating segmentation accuracy metrics on an external validation dataset. The additional training phase implemented for version 2.0 yielded an overall mean Dice similarity coefficient (DSC) improvement of 4% when looking at lesions of any size and class (range 0-13% for individual lesion classes). Mean absolute volume errors between automated and ground-truth segmentations also improved for most lesion types (2.94 vs. 1.55 mL for IPH, 18.44 vs. 16.33 mL for EAH, 0.74 vs. 0.80 for IVH, and 1.56 vs. 0.27 for perilesional edema using version 1.0 vs. 2.0, respectively). Overall, the performance of BLAST-CT on the PROTEST external validation dataset was comparable or better to the results obtained on our internal development sample (median DSC for all lesion classes was 0.60 [IQR 0.0-0.94] on the PROTEST images vs. 0.36 [IQR 0.0-0.63] on the CENTER-TBI development dataset). Mean volume differences between the ground-truth and predicted lesion maps were also comparable to the internal sample for most lesion subtypes, with the exception of EAH. We propose one of the first models capable of automated multiclass volumetric lesion segmentation in TBI to be trained and externally validated in a multicenter fashion. After optimizing our model using a large additional training sample from CENTER-TBI, we were able to achieve a level of performance comparable to other state-of-the-art methods. We also make this optimized model (BLAST-CT 2.0) publicly available to provide a robust and generalizable platform for application to large prospective TBI datasets and clinical trials.

INTRODUCTION: The decision to withdraw life sustaining treatment (WLST) for pediatric patients with severe traumatic brain injury (TBI) is challenging for clinicians and families with limited evidence quantifying existing practices. METHODS: This retrospective study utilized data collected from trauma centers through the American College of Surgeons Trauma Quality Improvement Program between 2017-2020. We included pediatric patients (<19 years) with severe TBI and a documented decision for WLST. We utilized a random intercept multilevel logistic regression model to quantify patient, injury and hospital characteristics influencing WLST. In order to quantify the impact of disparate WLST practices on mortality, we ranked centers by their conditional random intercept computed quartile-specific adjusted mortality. RESULTS: We identified a total of 9803 children with severe TBI treated across 515 trauma centers, of which 1003 (10.2%) had WLST. Patient-level factors associated with increased likelihood of WLST were young age (<3 years), higher severity intracranial and extracranial injuries and mechanism of injury related to firearms. Following adjustment for patient and hospital attributes, the median odds ratio was 1.54 (95% CI: 1.50-1.57), reflecting residual variation in WLST between centers. When centers were grouped into quartiles by their propensity for WLST, adjusted mortality was higher for fourth compared to first quartile centers (OR 1.67, 95% CI: 1.47-1.90). CONCLUSIONS: Several patient and injury factors were associated with WLST decision-making for pediatric patients with severe TBI. We also showed variation in WLST practices between trauma centers after adjustment for case mix; these differences were associated with differing quartile-specific adjusted mortality rates. Taken together, these findings highlight the presence of care provision inequities among North American trauma centers regarding WLST practice patterns for pediatric patients with severe TBI.

INTRODUCTION: Artificial intelligence (AI) model integration into clinical workflow offers potential to optimize decision-support for transfer of acute traumatic brain injury (TBI) patients to appropriate trauma centers. METHODS: We retrospectively identified TBI patients from 2005-2021 treated at a quaternary Canadian trauma center. We employed various modeling techniques including principal component analysis, three-dimensional convolutional neural networks, and a transformer-based approach using Vision Transformer (ViT) architecture. Model training, validation, and testing was performed using head CT scans with binary ground truth labels corresponding to whether the patient received neurosurgical intervention witin 72 hours. The finalized model, termed Automated Surgical Intervention Support Tool for TBI (ASIST-TBI), was then deployed in a simulated prospective fashion on consecutive TBI patients at our center between March 2021 - September 2022. RESULTS: A dataset of 2,806 trauma patients with acute head CT scans were divided into training, validation, and testing groups; the ViT model exhibited optimal performance. There was accurate prediction of requirement for neurosurgical intervention with an area under the receiver operating curve (AUC) of 0·92, accuracy of 0·87, sensitivity of 0·87, and specificity of 0·88 on the testing cohort. In simulated 18-month prospective deployment, an additional 612 consecutive scans were used to assess the performance of ASIST-TBI. Classification accuracy remained robust with AUC of 0·89, 0·85 sensitivity, 0·84 specificity, and 0·84 accuracy. We manually reviewed false positive and false negative cases to identify reasons for misclassification. CONCLUSIONS: We developed a novel deep learning model that accurately predicts requirement for acute neurosurgical intervention using unlabeled TBI scans. ASIST-TBI has potential application to optimize state-wide triage efficiency and care pathways for brain-injured patients.

INTRODUCTION: Withdrawal of life sustaining treatment (WLST) in severe traumatic brain injury (TBI) is complex with a paucity of standardized guidelines to inform practice patterns. We hypothesized presence of variability in WLST decision-making across centers, which could have important implications for equitable trauma care provision. METHODS: This retrospective study utilized data from adult severe TBI patients treated at trauma centers participating in the Trauma Quality Improvement Program between 2017-2020. Multivariable hierarchical logistic regression was used to adjust for patient, injury and hospital attributes influencing WLST. Residual between-center variability was characterized using the median odds ratio. Disparate WLST practices were further assessed by ranking centers' tendencies for WLST and assessing mortality between quartiles. RESULTS: We identified a total of 85511 subjects with severe TBI treated across 510 trauma centers, of which 20,300 (24%) had WLST. Patient-level factors associated with increased likelihood of WLST were advanced age, white race, treatment in a non-profit trauma center or self-pay/ Medicare insurance status (compared to private insurance). Black race was associated with reduced tendency for WLST. Higher severity intracranial and extracranial injuries also increased likelihood for WLST.After adjustment for patient and hospital attributes, the median odds ratio was 1.44 (1.40-1.48 95%CI), suggesting substantial residual variation in WLST between centers. When centers were grouped into quartiles by their propensity for WLST, there was a 1.37 (1.30-1.42 95%CI) increased adjusted odds of mortality between fourth versus first quartile centers. CONCLUSIONS: We highlighted the presence of disparate WLST practice patterns between trauma centers after adjustment for case-mix and hospital attributes. These findings highlight a strong need for examination of center-level factors contributing to heterogenous WLST practices to ultimately improve equity of care provision for severe TBI patients.