A06 Utilizing point-of-care informed feature organization and multimodality information integration to develop artificial intelligence systems for effective blood transfusion triage

Background Massive haemorrhage remains a leading cause of preventable mortality on the battlefield. 1 Timely and effective decision-making around blood transfusion across various points of care is complex and often suboptimal, leading to avoidable patient deterioration. Artificial intelligence and machine learning (AI/ML) approaches offer significant potential for improving triage accuracy and optimizing blood resource allocation. 2 3 This preliminary study introduces a point-of-care informed clinical feature organization and multimodality information integration framework to identify key predictive features and develop initial ML models. The overarching goal is to lay the foundation for an AI-assisted transfusion triage system that enhances both efficiency and outcomes. Methods A retrospective dataset from the Ontario Trauma Registry was utilized, comprising 73,117 trauma patients and 604 multimodal clinical features, stratified into pre-hospital (31 features) and in-hospital (104 features) phases of care. Two working, outcome-balanced datasets (each >3,000 samples) were created for separate phase-specific analyses. A comprehensive data pre-processing strategy was implemented. The CV-rRF-FS-SVM (cross validation with recursive random forest and support vector machine) algorithm was applied via NMT (Neural ML Tools) toolkit for feature selection, with transfusion status as the outcome variable. 4–6 SVM was then employed to develop preliminary transfusion triage models. Further, explainable AI method SHAP (SHapley Additive exPlanations) analysis was applied for better model and individual inference interpretation. Additionally, an AI-powered Large Language Model (LLM) framework, termed ‘DefencePT’, was established to analyse open-ended language data within the dataset, enabling topic extraction and more comprehensive data exploitation. Results In the pre-hospital phase, 10 selected critical predictors included heart rate at the scene and primary injury type. The SVM model achieved an ROC-AUC of 0.82 on a 20% holdout test set (figure 1A). For the in-hospital analysis, 21 key features were identified, including heart rate, temperature, injury type, and systolic blood pressure, with a resulting ROC-AUC of 0.88 (figure 1B). The SHAP results suggested that the preliminary models can be explained, on both the model- and individual-levels. The DefencePT LLM framework revealed previously unstructured insights, such as associations between injury types and age demographics, offering early evidence to better understand preventable haemorrhage cases and improve blood resource management strategies. Conclusions By implementing a point-of-care informed feature organization and multimodality integration strategy, our machine learning analysis has effectively identified key determinants in blood transfusion triage in both pre-hospital and in-hospital trauma care settings. This approach uncovers subtle patterns and trends that are not readily observable through traditional methods, offering potential to enhance blood transfusion protocols and practices. The LLM-powered framework demonstrated promise in extracting actionable insights from unstructured data, potentially improving triage model performance and informing strategies to reduce preventable battlefield mortality. Collectively, these preliminary findings underscore the transformative potential of AI/ML in advancing transfusion triage practices both on and off the battlefield. Abstract A06 Figure 1 ROC-AUC results on 20% holdout test data on the preliminary transfusion triage models: (A) pre-hospital model, (B) in-hospital model Conflict of Interest Authors declare no conflict of interest. References Alam HB, Burris D, DaCorta JA, Rhee P. Hemorrhage control in the battlefield: role of new hemostatic agents. Mil Med . 2005; 170 (1):63–9. Anstey C, Ullman D, Su L, Su C, Siniard C, Simmons S, et al . The practical use of artificial intelligence in transfusion medicine and apheresis. Transfus Apher Sci . 2024; 63 (6):104001. Peng HT, Siddiqui MM, Rhind SG, Zhang J, da Luz LT, Beckett A. Artificial intelligence and machine learning for hemorrhagic trauma care. Military Medical Research . 2023; 10 (1):6. Zhang J, Wong SM, Richardson JD, Jetly R, Dunkley BT. Predicting PTSD severity using longitudinal magnetoencephalography with a multi-step learning framework. Journal of neural engineering . 2020; 17 (6). Zhang J, Richardson JD, Dunkley BT. Classifying post-traumatic stress disorder using the magnetoencephalographic connectome and machine learning. Scientific Reports . 2020; 10 (1):5937. Zhang J, Hadj-Moussa H, Storey KB. Current progress of high-throughput microRNA differential expression analysis and random forest gene selection for model and non-model systems: an R implementation. Journal of integrative bioinformatics . 2016; 13 :306.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.