Diagnostic and prognostic applications of machine learning in paediatric traumatic brain injury: a systematic review of single and multimodal approaches.

Authors

Abstract

Paediatric traumatic brain injury (pTBI) is one of the leading causes of global childhood disability. While traditional tools like the Glasgow Coma Scale (GCS) are clinical staples, they often lack the precision required for individualized care. This systematic review evaluates machine learning (ML) and deep learning (DL) models for both the acute detection of injuries and the long-term prognostication of outcomes in pTBI. Following PRISMA guidelines and PROSPERO registration (CRD42024510419), a comprehensive search of five databases through May 2026 was conducted to include recent advancements in transformer architectures and gradient boosting. Inclusion criteria focused on paediatric patients from 0 to 18 years, using single or multimodal data (clinical, imaging, and biomarkers) for TBI identification or outcome prediction. Methodological quality was assessed using QUIPS and PROBAST. Twenty full-text studies (comprising 68,331 subjects) were analyzed. Models were stratified into acute diagnosis (e.g., skull fracture detection, CT-necessity triage) and outcome prognostication (e.g., mortality, 6-month functional recovery). Algorithms such as XGBoost, Random Forest, and CNNs showed potential to outperform traditional regression in specific scenarios, achieving AUROCs up to 0.98 for mortality. However, gains were marginal in low-risk triage, where ML models did not significantly surpass the "no-information rate." ML models show significant potential for enhancing pTBI care through improved risk stratification and automated imaging analysis. Future implementation requires a focus on model interpretability (e.g., SHAP values), addressing class imbalances, and conducting external multicenter validation to ensure regional generalizability. https://www.crd.york.ac.uk/PROSPERO/view/CRD42024510419, identifier CRD42024510419.

Tags

Topics