Sex-specific differences in mortality after adult traumatic brain injury (TBI) have produced inconsistent findings across the literature, even when studies adjust for age, injury severity, and related factors. The authors aimed to synthesize adjusted mortality estimates comparing males and females after adult TBI via systematic review and meta-analysis, and to use explainable machine learning to investigate study-level factors that might modify reported effect estimates.
The review searched eight databases from inception through August 25, 2025. Studies were eligible if they reported mortality estimates comparing sexes in adult TBI and provided estimates adjusted for at least age and injury severity. From 9,418 records identified, 21 studies met inclusion criteria. The assembled data set included 899,813 participants and 98,286 deaths.
Risk of bias for included studies was assessed using the Quality in Prognosis Studies (QUIPS) tool. The authors also documented study-level characteristics used in subsequent analytic steps and for machine-learning analyses. Details on data extraction, transformations for ML, and the minimal data set and code are reported in the manuscript and submitted appendices.
Male-to-female mortality effect estimates from included studies were pooled using random-effects meta-analysis to accommodate between-study heterogeneity. The meta-analytic pooling produced the following primary estimates:
The authors also reported pooled results by mortality timing or definition, noting consistent directionality favoring lower female mortality in several outcome groupings.
When analyses were restricted by outcome type, lower mortality among females remained evident for specific categories:
The authors conducted sensitivity analyses to test the robustness of pooled results; they report that the primary findings were consistent across those sensitivity analyses. Nonetheless, substantial residual heterogeneity remained across pooled analyses, indicating that other study-level factors explain variation between reported effect sizes.
To explore sources of between-study heterogeneity, the investigators applied three explainable machine learning algorithms to study-level data and effect estimates. These algorithms were used to identify variables that modify reported male-to-female mortality associations across studies. Explainable ML methods highlighted the following important modifiers of effect estimates:
The ML findings were reported as consistent across sensitivity analyses, suggesting that demographic and methodological study characteristics contribute meaningfully to observed heterogeneity in sex-specific mortality estimates after TBI.
The pooled meta-analytic estimates suggest that female sex is associated with lower mortality after adult TBI, both in odds and hazard metric summaries. However, the presence of substantial residual heterogeneity indicates that pooled effect estimates mask important between-study differences.
Explainable ML results implicate demographic factors (notably mean age and sex composition) and methodological quality as contributors to variation in reported associations. The authors emphasize that these sources of heterogeneity are relevant to prognostic research and the creation, validation, and interpretation of data-driven clinical decision-support or digital prognostic tools for TBI. When developing and validating such tools, investigators should account for population characteristics and study context to improve generalizability and reduce bias.
Strengths reported by the authors include a comprehensive search across multiple databases, inclusion of studies with adjusted estimates for at least age and injury severity, and integration of explainable ML to probe modifiers of study-level effect estimates. The authors assessed study risk of bias using the Quality in Prognosis Studies tool and conducted sensitivity analyses to evaluate robustness.
Limitations described in the source include substantial residual between-study heterogeneity that was not fully explained by pooled analyses. The original report indicates the findings are based on published literature; no new human participants were recruited. The authors state that the minimal data set and accompanying code are available in the manuscript and appendices. Specific numerical or study-level details beyond those reported were not provided in the source text.
Across 21 studies including nearly 900,000 participants, pooled adjusted estimates indicate lower mortality among females after adult TBI. Nonetheless, heterogeneity across studies remains substantial. Explainable machine learning identified mean age, study sex composition, and study quality as important modifiers of reported effect estimates. The authors conclude that demographic and methodological context must be considered when developing and applying prognostic and decision-support tools for TBI. The review is registered in PROSPERO (CRD420251252136). The authors declared no competing interests and acknowledged funding sources as reported in the manuscript.