Development, validation, and usability evaluation of machine learning algorithms for predicting personalized red...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Surgical Blood Management

Background:

• Accurate red blood cell (RBC) preparation is vital for perioperative workflow and patient safety, especially in thoracic surgery (TS) due to high transfusion variability.
• Precise prediction of individual RBC requirements for TS patients is increasingly critical.
• This study addresses the need for personalized RBC predictions in TS.

Purpose of the Study:

• To develop and validate a machine learning algorithm for personalized RBC predictions in thoracic surgery patients.
• To assess the usability of a clinical decision support system (CDSS) integrating this AI model.
• To enhance the precision of blood preparation for thoracic surgeries.

Main Methods:

• Developed and compared traditional statistical and machine learning models using data from 7,843 adult TS patients (2016-2021).
• Selected the extreme gradient boosting model, termed pMSBOS-Thoracic Surgery (pMSBOS-TS), based on performance metrics (RMSE, adjusted R²).
• Assessed the usability of the pMSBOS-TS CDSS with 60 clinicians using the System Usability Scale (SUS).

Main Results:

• The pMSBOS-TS algorithm demonstrated significantly lower RMSE and higher adjusted R² compared to the standard Maximum Surgical Blood Ordering Schedule (MSBOS).
• The AI model predicted RBC requirements more accurately, leading to approximately 200 fewer RBC units prepared compared to MSBOS.
• The pMSBOS-TS CDSS achieved a SUS score of 72.5, indicating good usability and acceptability among clinicians.

Conclusions:

• Successfully developed the pMSBOS-TS, an AI-driven tool for personalized RBC transfusion requirement prediction in perioperative thoracic surgery.
• The pMSBOS-TS offers improved accuracy and efficiency in blood preparation for TS.
• The developed CDSS shows good clinical usability, supporting its integration into surgical workflows.