Comparative predictive performance of computed tomography scoring systems in traumatic brain injury: a systematic review, Bayesian comparison, and meta-analysis.

Various computed tomography (CT)-based Traumatic brain injury (TBI) scoring systems, including Marshall, Rotterdam, Helsinki, Stockholm, and the more recent NeuroImaging Radiological Interpretation System (NIRIS), have been specifically developed to estimate prognosis based on radiological features. We aimed to comprehensively compare the CT scoring systems’ predictive accuracy across diverse clinical outcomes. Bayesian multilevel logistic regression was applied to assess the predictive performance of scoring systems for in-hospital, 1-month, 6-month, and 12-month mortality, as well as unfavorable outcomes. Moreover, area under the curve (AUC) values were pooled across all outcomes using random-effects models. Pooled meta-analysis on AUC aimed to evaluate the predictive performance of each CT scoring system across different outcomes. A total of 32 studies involving 23,529 patients were included. In 32 included studies, the Rotterdam scoring system was the most frequently used in 26 studies (81%). Followed by Marshall in 17 studies (53%), Helsinki in 10 studies (31%), Stockholm in four studies (12%), and the NIRIS system in only one study (3%). The pooled meta-analysis on AUC across all outcomes for the Rotterdam score using AUC was 0.774 (0.751–0.797). The Marshall score revealed a moderate predictive AUC of 0.695 (0.660–0.729). The Helsinki score had a pooled AUC of 0.772 (0.746–0.798). The Stockholm score demonstrated a relatively higher accuracy with an AUC of 0.778 (0.739–0.818). The Bayesian comparative analysis indicated that the Helsinki score has relatively higher accuracy for 6-month mortality (AUC: 0.838; 0.673–0.952) and 6-month unfavorable outcomes (AUC: 0.785; 0.666–0.904). The Stockholm score showed the highest accuracy for 12-month unfavorable outcomes (AUC: 0.687; 0.577–0.780) and 12-month mortality (AUC: 0.614; 0.430–0.769). For shorter-term outcomes, the Marshall score had relatively higher accuracy for in-hospital mortality (AUC: 0.777; 0.593–0.901), while the Rotterdam score demonstrated relatively better performance for 1-month mortality (AUC: 0.753; 0.567–0.867). While the overlapping credible intervals suggest that no single system consistently outperformed the others, Marshall and Rotterdam scores showed relatively higher predictive accuracy for short-term outcomes, while Helsinki and Stockholm performed relatively better for longer-term prognostication. Evidence for NIRIS remains limited. We recommend Marshall/Rotterdam for acute-phase risk stratification and Helsinki/Stockholm for longer-term prediction, with future research focusing on multimodal models integrating clinical, imaging, and biomarker data.