Abstract
Over half of all deaths after traumatic brain injury (TBI) follow the decision to withdraw life-sustaining therapy (WLST). Despite recent improvements in TBI mortality, rates of WLST have remained unchanged, potentially reflecting outdated prognostic misconceptions. Determinants of the decision to WLST are multifaceted and complex, spanning just beyond clinical prognosis, which is reflected in the significant variability in WLST rates between centers. We aimed to create a machine learning (ML) model that could accurately predict the decision to WLST and hypothesized that facility WLST rate would emerge as a highly impactful WLST determinant.
This observational study analyzed data from the American College of Surgeons Trauma Quality Improvement Project National Trauma Databank (2017-2021). Patients with severe TBI, defined by a maximum Abbreviated Injury Scale-Head ≥1 and presenting Glasgow Coma Scale (GCS) <9, were included. Burns, unspecified injury mechanisms, emergency room discharges, deaths, transfers, and patients with missing WLST data were excluded. ML models were developed to predict WLST using variables available at different time points. The performance of each model in predicting WLST was optimized for area under the receiver operating curve (AUROC). The most impactful determinants of WLST were assessed using Shapley additive explanation scores.
Of 5,481,046 patients, 155,639 met inclusion criteria, with 32,385 (20.8%) undergoing WLST. The mean age was 43 ± 22 years, 26.5% of patients were female, and the median time to WLST was 46.4 hours. The AUROC of 0.875 (95% CI 0.871-0.879) in the admission model improved to 0.896 (95% CI 0.892-0.900) in the total length-of-stay model. Age, highest emergency department GCS, and facility WLST rate were the most important factors in prediction of WLST.
In this study of using ML to predict WLST after severe TBI, our models reliably predict the decision to WLST. We found that institutional withdrawal culture is a strong independent determinant of WLST, irrespective of clinical condition. As TBI care improves, our findings underscore the importance of refining prognosticating tools to prevent premature WLST decisions which may be influenced by biases associated with self-fulfilling prophecies and institutional practice patterns.
