Article révisé par les pairs
Résumé : Objectives Liver transplantation is a complex procedure frequently requiring transfusion of blood products to manage coagulopathy and haemorrhage. This study aimed to develop machine learning models to predict the biological effects of blood product transfusions, assisting clinicians in selecting optimal therapeutic combinations. Methods Using data from two cohorts over 20 years from two academic hospitals, 10 supervised machine learning models were trained and validated on four biomarkers: fibrinogen, haemoglobin, prothrombin time and activated partial thromboplastin time ratio. Models were evaluated using R², root mean squared error and SD metrics, with external validation performed on the second cohort. Results The results indicated that while certain models, such as the stack model for late fibrinogen (R²=0.63) or the extra trees model for late prothrombin time (R²=0.66), demonstrated promising predictive capacity, the overall external validation performance was suboptimal. Despite the use of a large healthcare database, a rigorous statistical methodology and an academic machine learning methodology, most models showed limited generalisability (R² < 0.5). Discussion Key limitations included the small dataset size relative to machine learning requirements, lack of advanced haemostatic parameters (eg, ROtational ThromboElastoMetry (ROTEM) or Thromboelastography (TEG)) and the variability introduced by evolving surgical practices over the 20-year study period. Despite these limitations, this study provides a reproducible framework for evaluating transfusion efficacy, supported by openly shared Python code and the application of Taylor diagrams for model evaluation. Conclusion While our models are unsuitable for routine clinical use, they highlight the potential of machine learning in transfusion medicine. Future work should focus on integrating larger datasets, advanced biomarkers and real-time data.

Documents en relation

DI-fusion