Personalized Patient Blood Management for oncological patients with machine learning

Machine Learning (ML) techniques are increasingly used to enhance the safety and efficacy of blood transfusions, supporting Patient Blood Management (PBM) strategies. This study aimed to develop and validate a predictive model for the risk of intraoperative transfusions in oncological patients undergoing complex surgeries, with the aim to improve transfusion appropriateness and enable personalized preoperative planning. Clinical and laboratory data from a cohort of oncological surgery patients were analyzed using a Catboost model. The model achieved Positive Predictive Values (PPVs) of 0.7 and 0.6, and Negative Predictive Values (NPVs) of 0.9 and 0.8 for the training and testing datasets, respectively, with AUC scores of 0.89 for training and 0.8 for testing. Explainability analysis identified clinical and laboratory predictors of transfusion risk, such as hematocrit, hemoglobin, various blood counts, and tumor sites. 
The model demonstrated significant predictive capability for assessing intraoperative transfusion risk, which could help minimize transfusion-related complications and improve surgical outcomes for oncological patients.