Development and validation of a machine learning prediction model for perioperative red blood cell transfusions in
Qian Li1, Hong Lv1, Yuye Chen1
1Department of Anesthesiology, Fuwai Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College/National Center for Cardiovascular Diseases, Beijing 100037, China.
Objective:
Several machine learning (ML) models have been used in perioperative red blood cell (RBC) transfusion risk for cardiac surgery with limited generalizability and no external validation. Hence, we sought to develop and comprehensively externally validate a ML model in a large dataset to estimate RBC transfusion in cardiac surgery with cardiopulmonary bypass (CPB).
Design:
A retrospective analysis of a multicenter clinical trial (NCT03782350).
Patients:
The study patients who underwent cardiac surgery with CPB came from four cardiac centers in China and Medical Information Mart for Intensive Cared (MIMIC-IV) dataset.
Measurements:
Data from Fuwai Hospital were used to develop an individualized prediction model for RBC transfusion. The model was externally validated in the data from three other centers and MIMIC-IV dataset. Twelve models were constructed.
Main Results:
A total of 11,201 eligible patients were included in the model development (2420 in Fuwai Hospital) and external validation (563 in the other three centers and 8218 in the MIMIC-IV dataset). A significant difference was observed between the Logistic Regression and CatboostClassifier (0.72 Vs. 0.74, P = 0.031) or RandomForestClassifier (0.72 Vs. 0.75 p = 0.012) in the external validation and MIMIV-IV datasets (age ≤ 70:0.63 Vs. 0.71, p < 0.001; age > 70:0.63 Vs. 0.70, 0.63 Vs. 0.71, p < 0.001). The CatboostClassifier and RandomForestClassifier model was comparable in development (0.83 Vs. 0.82, p = 0.419), external (0.74 Vs. 0.75, p = 0.268), and MIMIC-IV datasets (age ≤ 70: 0.71 Vs. 0.71, p = 0.574; age > 70: 0.70 Vs. 0.71, p = 0.981). Of note, they outperformed other ML models with excellent discrimination and calibration. The CatboostClassifier and RandomForestClassifier models achieved higher area under precision-recall curve and lower brier loss score in validation and MIMIC-IV datasets. Additionally, we confirmed that low preoperative hemoglobin, low body mass index, old age, and female sex increased the risk of RBC transfusion.
Conclusions:
In our study, enrolling a broad range of cardiovascular surgeries with CPB and utilizing a restrictive RBC transfusion strategy, robustly validates the generalizability of ML algorithms for predicting RBC transfusion risk. Notably, the CatboostClassifier and RandomForestClassifier exhibit strong external clinical applicability, underscoring their potential for widespread adoption. This study provides compelling evidence supporting the efficacy and practical value of ML-based approaches in enhancing transfusion risk prediction in clinical practice.
