COMPARATIVE ANALYSIS OF MACHINE LEARNING VERSUS CONVENTIONAL MODELS FOR PREDICTING ERYTHROCYTE NEED IN CARDIOVASCULAR SURGERY

OBJECTIVE: Excessive bleeding during cardiovascular surgery poses significant risks to patients, necessitating accurate prediction of erythrocyte suspension (ES) requirements for optimal perioperative management. This study aims to compare the efficacy of machine learning (ML) algorithms against conventional scorings in predicting ES needs during coronary artery bypass grafting (CABG) surgeries. MATERIAL AND METHODS: The ML algorithm, developed using patient demographic, preoperative, and surgical data, was compared with six established bleeding prediction scores. ES need predictions were analyzed using logistic regression-based neural networks. RESULTS: A total of 430 patients undergoing isolated CABG surgeries were included in the analysis. ML algorithms demonstrated comparable predictive power to conventional scoring systems, with an accuracy of 75.5% for predicting any ES need and 93.8% for predicting the need for >2 units of ES. Among traditional scoring systems, the TRUST score exhibited the highest predictive ability, followed by the TRACK and WILL BLEED scores. CONCLUSIONS: The ML algorithm showed potential for local adaptation and self-improvement, enhancing its utility in perioperative planning. Our findings underscore the feasibility of ML algorithms in optimizing blood product utilization in cardiovascular surgery. Further research is warranted to explore the scalability and generalizability of these models across diverse clinical settings. ML predictions for ES transfusion readiness in patients prior to CABG surgery demonstrate a high degree of accuracy and comparable performance to traditional scoring systems. The ability of ML algorithms to reinforce themselves with local data and adapt to changing clinical contexts highlights their potential for enhancing predictive accuracy over time.