Development of a machine learning model to predict intraoperative transfusion and guide type and screen ordering

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Transfusion Medicine

Background:

• Optimizing blood product utilization is crucial in healthcare.
• Preoperative prediction of red blood cell (RBC) transfusion needs can improve efficiency.
• Current methods for ordering type and screens may lead to unnecessary testing.

Purpose of the Study:

• To develop and validate a machine learning algorithm for predicting intraoperative RBC transfusion.
• To utilize preoperative variables from electronic medical records for prediction.
• To guide and improve preoperative type and screen ordering practices.

Main Methods:

• Retrospective analysis of single-center hospital data (2019-2022).
• Development and comparison of seven machine learning algorithms.
• Model training on 2019-2021 data, with thresholds set to institutional sensitivity (93%).
• External validation on 2022 data, requiring sensitivity >90% for comparison.

Main Results:

• The LightGBM model achieved 76.1% overall accuracy with 91.2% sensitivity in predicting intraoperative transfusions.
• The model demonstrated external temporal validity on 2022 data.
• Implementing the LightGBM model could improve the type and screen to transfusion ratio from 8.4 to 5.1.

Conclusions:

• Machine learning models can effectively predict intraoperative RBC transfusions from preoperative data.
• Integration into electronic health records can enhance preoperative type and screen ordering.
• This approach holds potential for optimizing blood management and patient care.