Machine learning to predict haemorrhage after injury: So many models, so little dynamism

Abstract

Accurately predicting the need for blood transfusion in bleeding patients remains a critical challenge in emergency care. Machine learning (ML) models show promise for improving decision support in these scenarios, but a gap remains between research and practical application. Existing models frequently overlook the dynamic nature of clinical data, hindering their ability to provide accurate predictions for blood transfusion needs in emergency settings. We conducted a scoping review to examine ML models that integrate time-varying variables to predict blood transfusion needs in trauma patients. We discuss challenges in data collection, particularly the limitations of electronic health records (EHRs) in capturing high-quality time-series data and emphasise the need for explainable artificial intelligence (AI). We suggest future directions for research that include advancing computational approaches, improving data collection, and enhancing the interpretability of ML models to ensure their clinical relevance and utility.