Personalized Prognostication in Paediatric TBI: A Systematic Review of Machine Learning Approaches with Multimodal Data

Abstract

Background: Paediatric traumatic brain injury (pTBI) is a major global public health concern, contributing substantially to childhood disability and mortality. Traditional prognostic tools, such as the Glasgow Coma Scale (GCS), lack specificity and sensitivity, limiting their accuracy in predicting outcomes. This systematic review evaluated the use of machine learning (ML) models that integrate multimodal data to improve prognostication in pTBI. Methods: his systematic review was registered with PROSPERO and conducted according to PRISMA guidelines. We systematically searched MEDLINE, PubMed, Embase, Scopus, Web of Science, and Cochrane for studies published in the past decade that applied ML to predict clinical outcomes in patients under 18 years with pTBI using multimodal data. Study quality was assessed using the QUIPS tool. Results: Thirteen studies met inclusion criteria. ML algorithms, including support vector machines, random forest, artificial neural networks, CatBoost, and others, were applied to multimodal data incorporating clinical, radiological, neurocognitive, laboratory, and administrative variables. Most studies reported that ML models outperformed traditional methods, demonstrating higher sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC). However, performance varied across studies, and some showed only modest improvements over conventional approaches. Lack of standardized reporting and limited assessment of clinical significance were common limitations. Conclusion: ML models leveraging multimodal data show promise for improving prognostic accuracy in pTBI. Future research should address methodological heterogeneity, enhance interpretability and reproducibility, minimize false negatives, and adopt standardized reporting to support clinical implementation.