Abstract
Traumatic brain injury (TBI) remains a leading cause of death and disability worldwide, with outcomes that are highly heterogeneous and difficult to predict using conventional clinical tools. Artificial intelligence (AI) has emerged as a promising approach to enhance diagnosis, prognostication, and management of TBI across diverse care settings. Following PRISMA guidelines, a comprehensive search of PubMed, Scopus, Web of Science, and Cochrane CENTRAL was conducted from inception to September 2025. Eligible studies applied AI or machine learning techniques to TBI populations for diagnosis, outcome prediction, monitoring, or screening, with data extracted on study design, population characteristics, data sources, model architecture, comparators, and performance metrics. Methodological quality and risk of bias were assessed using the PROBAST tool. From 10,710 records identified, 12 studies met the inclusion criteria. Prognostic models across emergency triage, intensive care unit, and registry datasets achieved AUCs (area under the curve) between 0.81 and 0.93, with simple logistic regression models often performing comparably to more complex machine learning methods. Imaging-based approaches, including convolutional neural networks for CT segmentation, improved prediction of therapeutic intensity but were constrained by small sample sizes and inconsistent segmentation quality. Unsupervised clustering revealed clinically meaningful phenotypes, though generalizability was variable. Pediatric applications were exploratory, with small cohorts prone to overfitting. Screening models for mild TBI demonstrated overall accuracies of 0.80-0.86 but only modest sensitivity (~0.70) for CT-positive cases. Across domains, interpretability strategies such as SHAP values and fuzzy rules showed promise; however, many studies were limited by inadequate external validation, incomplete handling of missing data, and small or retrospective single-center designs. Overall, AI applications in TBI demonstrate strong potential for improving prognostication, imaging analysis, and patient stratification, yet their clinical translation remains constrained by methodological shortcomings. Parsimonious models leveraging readily available clinical variables frequently rival more complex approaches, underscoring the importance of task-specific model selection. Future research should prioritize multicenter prospective datasets, external validation, calibration assessment, and evaluation of clinical impact to enable reliable integration of AI into TBI care.