Premium
Machine learning for predicting preoperative red blood cell demand
Author(s) -
Feng Yannan,
Xu Zhenhua,
Sun Xiaolin,
Wang Deqing,
Yu Yang
Publication year - 2021
Publication title -
transfusion medicine
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.471
H-Index - 59
eISSN - 1365-3148
pISSN - 0958-7578
DOI - 10.1111/tme.12794
Subject(s) - medicine , perioperative , machine learning , artificial intelligence , surgery , computer science
Background The paucity of accurate quantitative standards for determining the quantity of red blood cells (RBCs) needed for perioperative patients and the predominant application of the “preoperative hemoglobin + surgery type” empirical decision‐making model have led to widespread RBC application problems. Objective The mathematical model of preoperative variables constructed by machine learning (ML) methods can help doctors decide preoperative RBC applications. Methods We retrospectively analysed 130 996 records of patients who received surgery in our hospital from January 2011 to June 2017. Through the analysis of multiple preoperative parameters that may affect the RBC transfusion volume, we used ML algorithms to build up the artificial intelligence (AI) model to predict the accurate RBC demand quantity and compared each result with those predicted by clinicians. Results Among the seven ML algorithms, the light gradient boosting machine (Lightgbm) algorithm was the best. The AI model predicted whether the patients needed RBC transfusion, and the area under curve (AUC) was 0.908 (95% CI 0.907–0.913). The AI model was more accurate than doctors in predicting RBC of 0, 2, and 4 units (85% data), with RMSEs of 1.61 vs. 2.15, 1.06 vs. 1.21, and 1.46 vs. 1.68, respectively. However, the AI model was not better than doctors in 1, 3, 5–6, 7–8, and 9–10 units (15% data), with RMSEs of 0.92 vs. 0.89, 0.92 vs. 0.89, 2.73 vs. 1.94, 4.53 vs. 3.92, and 6.26 vs. 5.08, respectively. Conclusion Through the comparison of seven ML methods, the Lightgbm algorithm‐based model is more accurate than clinician experience‐based in predicting preoperative RBC transfusion, which reduces the risk of untimely blood supply caused by insufficient preoperative blood preparation, and reduces the unnecessary cost of blood compatibility testing caused by excessive preoperative blood preparation.