Scholarlink

Background <p>Postoperative blood transfusion remains a major challenge in joint arthroplasty, with substantial variability in clinical decision-making. Therefore, this study aimed to develop and validate machine learning models for early postoperative transfusion risk assessment to support early postoperative decision-making, with the clinical objective of minimizing unnecessary transfusions and improving perioperative blood resource utilization after joint arthroplasty.</p> Methods <p>We retrospectively analyzed 1,954 patients who underwent primary or revision hip or knee arthroplasty between June 2019 and December 2024. Patients were categorized according to whether they received allogeneic transfusion within 72&#xa0;h postoperatively. The dataset was randomly divided into training (70%) and validation (30%) cohorts. Logistic regression (LR), random forest (RF), and support vector machine (SVM) models were developed based on 15 key variables identified through LASSO regression. Model discrimination, calibration, and clinical utility were evaluated using ROC curves, Hosmer–Lemeshow tests, Brier scores, calibration slopes/intercepts, and decision curve analysis (DCA). SHAP analysis was used to interpret variable importance.</p> Results <p>Among all patients, 172 (8.8%) received transfusions. All three models showed good discrimination (AUC &gt; 0.75). The RF model achieved the best overall performance (training AUC = 1.000; validation AUC = 0.808, sensitivity = 87.5%, specificity = 59.0%; Brier score = 0.0651). The LR model also demonstrated stable calibration (Hosmer–Lemeshow <i>p</i> = 0.097, Brier = 0.0679), while the SVM model achieved moderate discrimination (validation AUC = 0.776) but weaker calibration. DCA confirmed that the RF model provided the greatest net clinical benefit within a threshold range of 0.1–0.75. SHAP analysis of RF model showed that postoperative hemoglobin, preoperative hemoglobin and operation time were the most important predictors.</p> Conclusion <p>Postoperative and preoperative haemoglobin levels, along with operative duration, were important predictors of early postoperative transfusion. The RF model demonstrated good predictive performance and clinical value, supporting its potential use for individualized transfusion risk assessment in patients undergoing joint arthroplasty.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Risk assessment and prediction of early blood transfusion after joint replacement surgery: a clinical decision support model based on machine learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月27日

查看全文
  • Tianyou Xing,
  • Jincai Duan,
  • Tianjie Xiao,
  • Zhihui Wang,
  • Huigeng Zhao,
  • Wei Qin,
  • Di Wu,
  • Changjiang Shi,
  • Yuanliang Du

摘要

Background

Postoperative blood transfusion remains a major challenge in joint arthroplasty, with substantial variability in clinical decision-making. Therefore, this study aimed to develop and validate machine learning models for early postoperative transfusion risk assessment to support early postoperative decision-making, with the clinical objective of minimizing unnecessary transfusions and improving perioperative blood resource utilization after joint arthroplasty.

Methods

We retrospectively analyzed 1,954 patients who underwent primary or revision hip or knee arthroplasty between June 2019 and December 2024. Patients were categorized according to whether they received allogeneic transfusion within 72 h postoperatively. The dataset was randomly divided into training (70%) and validation (30%) cohorts. Logistic regression (LR), random forest (RF), and support vector machine (SVM) models were developed based on 15 key variables identified through LASSO regression. Model discrimination, calibration, and clinical utility were evaluated using ROC curves, Hosmer–Lemeshow tests, Brier scores, calibration slopes/intercepts, and decision curve analysis (DCA). SHAP analysis was used to interpret variable importance.

Results

Among all patients, 172 (8.8%) received transfusions. All three models showed good discrimination (AUC > 0.75). The RF model achieved the best overall performance (training AUC = 1.000; validation AUC = 0.808, sensitivity = 87.5%, specificity = 59.0%; Brier score = 0.0651). The LR model also demonstrated stable calibration (Hosmer–Lemeshow p = 0.097, Brier = 0.0679), while the SVM model achieved moderate discrimination (validation AUC = 0.776) but weaker calibration. DCA confirmed that the RF model provided the greatest net clinical benefit within a threshold range of 0.1–0.75. SHAP analysis of RF model showed that postoperative hemoglobin, preoperative hemoglobin and operation time were the most important predictors.

Conclusion

Postoperative and preoperative haemoglobin levels, along with operative duration, were important predictors of early postoperative transfusion. The RF model demonstrated good predictive performance and clinical value, supporting its potential use for individualized transfusion risk assessment in patients undergoing joint arthroplasty.