AUTHOR=Yang Bo , Gao Lin , Wang Xingang , Wei Jianmin , Xia Bin , Liu Xiangwei , Zheng Peng TITLE=Application of supervised machine learning algorithms to predict the risk of hidden blood loss during the perioperative period in thoracolumbar burst fracture patients complicated with neurological compromise JOURNAL=Frontiers in Public Health VOLUME=10 YEAR=2022 URL=https://www.frontiersin.org/journals/public-health/articles/10.3389/fpubh.2022.969919 DOI=10.3389/fpubh.2022.969919 ISSN=2296-2565 ABSTRACT=Background

Machine learning (ML) is a type of artificial intelligence (AI) and has been utilized in clinical research and practice to construct high-performing prediction models. Hidden blood loss (HBL) is prevalent during the perioperative period of spinal treatment and might result in a poor prognosis. The aim of this study was to develop a ML-based model for identifying perioperative HBL-related risk factors in patients with thoracolumbar burst fracture (TBF).

Methods

In this study, single-central TBF patients were chosen. The medical information on patients, including clinical characteristics, laboratory indicators, and surgery-related parameters, was extracted. After comparing various ML model algorithms, we selected the best model with high performance. The model was validated using the internal validation set before performing recursive feature elimination (RFE) to determine the importance of HBL-related risk factors. The area under the receiver operating characteristic (AUC) curve, accuracy (ACC), sensitivity, and specificity were reported as critical model measures for evaluating predictive performance.

Results

In this study, 62 (38.5%) of the 161 TBF patients were positive for HBL. There was a significant statistical difference in age, body mass index (BMI), diabetes, hypertension, Beta (percentage of vertebral restoration), duration of operation, and other pre-operative laboratory indicators between the HBL-positive and HBL-negative groups. Nine ML-based models were built and validated, with the Random Forest model having the greatest AUC in both the training set (0.905) and internal validation set (0.864). Furthermore, following RFE, age, duration of operation, Beta, pre-operative fibrinogen (Fib), and activated partial thromboplastin time (APTT) were identified as the five main important risk factors in patients with TBF during the perioperative period.

Conclusion

In this study, we built and validated ML algorithms for an individualized prediction of HBL-related risk factors in the perioperative period of TBF. The importance of HBL-related risk factors could be determined, which contributes to clinicians' decision-making and improves perioperative management.