Deep learning-based survival outcomes of REBOA vs resuscitative thoracotomy in trauma: a nationwide cohort study in South Korea

Resuscitative endovascular balloon occlusion of the aorta (REBOA) has been used for trauma resuscitation and hemorrhage control, traditionally managed with resuscitative thoracotomy (RT). Despite growing evidence supporting REBOA, most studies have employed Kaplan–Meier (KM) analysis due to limited covariate data[1]. Therefore, using a nationwide, multi-center trauma cohort in South Korea, we aimed to compare survival between REBOA and RT, and evaluate whether a deep learning model could improve risk adjustment over KM by incorporating detailed clinical covariates. We utilized data from the Korean Trauma Data Bank, a nationwide registry capturing all consecutive trauma cases by integrating data from all 19 designated major trauma centers in South Korea from 1 January 2017 to 31 December 2022 (dataset No. K2024-01-2-03-29). This study was approved by the Institutional Review Board of Cheju Halla General Hospital (CHH-2024-L07), with a waiver of informed consent due to the secondary use of administrative data. Among 227 567 patients initially enrolled, 224 patients who received RT or REBOA were included (Supplemental Digital Content Table S1, available at: https://links.lww.com/JS9/G811). To compare survival outcomes, we constructed KM survival curves and developed a deep learning-based survival model following the TRIPOD + AI and TITAN 2025 guidelines on the use of AI in clinical research[2]. We employed neural multi-task logistic regression (N-MTLR), incorporating demographics and prehospital or in-hospital vital signs as covariates (Supplemental Digital Content Table S2, available at: https://links.lww.com/JS9/G811). Additionally, as a sensitivity analysis, we applied least absolute shrinkage and selection operator (LASSO) regularization to preselect informative covariates and retrained the N-MTLR model to assess the robustness of our survival estimates (Supplemental Digital Content Table S3, available at: https://links.lww.com/JS9/G811). This approach enabled multivariable-adjusted, nonlinear survival prediction and more accurate modeling of individual risk trajectories[3]. To enhance robustness, 95% confidence intervals (CIs) were estimated using a bootstrap resampling procedure. The predictive accuracy of KM and N-MTLR was evaluated using the integrated Brier score (IBS; Supplemental Digital Content Figure S1, available at: https://links.lww.com/JS9/G811), and interpretability was assessed with Shapley Additive exPlanations analysis (Supplemental Digital Content Figure S2, available at: https://links.lww.com/JS9/G811). Based on the better-performing model, subgroup analyses were conducted by (1) the presence of prehospital emergency care and (2) injury location (urban vs rural). Among the 224 patients, 162 underwent RT and 62 received REBOA [mean age: 48.9 (SD, 3.6) years for RT; 51.0 (4.0) years for REBOA; Table 1]. Mean survival probabilities were higher in the REBOA group than in the RT group by both KM [0.20 (95% CI, 0.10–0.31) versus 0.09 (0.05–0.14); log-rank P-value