Machine Learning for Preoperative Transfusion Risk Prediction in Total Knee and Hip Arthroplasty
Abstract
Background: Unnecessary blood preparation before total knee arthroplasty (TKA) and total hip arthroplasty (THA) increases healthcare costs and resource burden. Accurate preoperative transfusion risk prediction can support clinical decision-making and optimise blood bank management.
Methods: We conducted a retrospective cohort study using real-world electronic health records from a tertiary medical centre in Taiwan (2020–2025). A total of 6,089 patients undergoing TKA (n = 4,311) or THA (n = 1,778) were included. Twenty-eight preoperative features were extracted, including haematological indices, demographic variables, comorbidities, and medication use. Four machine learning models (Logistic Regression, Random Forest, XGBoost, LightGBM) were trained using 5-fold stratified cross-validation on an 80% development set and evaluated on a 20% holdout set. Model calibration was assessed using the Hosmer-Lemeshow test. SHAP values were used to interpret feature contributions globally and by surgical subgroup.
Results: The overall transfusion rate was 18.7% (TKA: 15.3%; THA: 27.1%). LightGBM achieved the highest holdout AUROC (0.751; 95% CI: 0.714–0.787), with good calibration (Hosmer-Lemeshow p = 0.297), Brier Score of 0.128, and NPV of 0.891 at the Youden Index threshold. SHAP analysis identified preoperative haemoglobin (Hb), surgery type, and tranexamic acid (TXA) use as the top predictors. A non-linear relationship was observed: transfusion risk increased sharply when preop Hb fell below 12 g/dL.
Conclusions: An interpretable LightGBM model demonstrated reliable discrimination and calibration for perioperative transfusion prediction. The high NPV suggests this model can safely identify low-risk patients, potentially reducing unnecessary blood preparation in orthopaedic surgical settings.
Key messages: • A LightGBM model predicts perioperative transfusion risk in TKA/THA with an AUROC of 0.751 and NPV of 0.891, enabling safer reduction of unnecessary blood preparation.
• Preoperative haemoglobin, surgery type, and TXA use are the strongest predictors of transfusion; risk rises sharply when Hb falls below 12 g/dL.
Related
- Explainable machine learning for preoperative prediction of acute kidney injury after intertrochanteric fracture surgery in older adults
- Development and validation of machine learning models for preoperative prediction of perioperative outcomes in patients with thymic epithelial tumors
- Development and validation of a machine learning model for preoperative prediction of intraoperative hypothermia in gynecological laparoscopic surgery: a two-center cohort study
- PREOPERATIVE HEMOGLOBIN AS A PREDICTOR OF TRANSFUSION REQUIREMENT IN PATIENTS UNDERGOING ELECTIVE HIP ARTHROPLASTY
- Preoperative Prediction of Difficult Laparoscopic Cholecystectomy: Evolution, Validation, and Emerging Machine-Learning Approaches-A Scoping Review.
BibTeX
@article{tien2026machine,
title = {{Machine Learning for Preoperative Transfusion Risk Prediction in Total Knee and Hip Arthroplasty}},
author = {TT Tien and CY Chen and HE Tzeng and JH Lin and KM Lu},
journal = {The European Journal of Public Health},
year = {2026},
month = oct,
doi = {10.1093/eurpub/ckag169.1207},
url = {https://doi.org/10.1093/eurpub/ckag169.1207},
note = {CC BY-NC 4.0},
}
RIS
TY - JOUR
TI - Machine Learning for Preoperative Transfusion Risk Prediction in Total Knee and Hip Arthroplasty
AU - TT Tien
AU - CY Chen
AU - HE Tzeng
AU - JH Lin
AU - KM Lu
PY - 2026
DA - 2026/10/01
JO - The European Journal of Public Health
DO - 10.1093/eurpub/ckag169.1207
UR - https://doi.org/10.1093/eurpub/ckag169.1207
AB - Background: Unnecessary blood preparation before total knee arthroplasty (TKA) and total hip arthroplasty (THA) increases healthcare costs and resource burden. Accurate preoperative transfusion risk prediction can support clinical decision-making and optimise blood bank management. Methods: We conducted a retrospective cohort study using real-world electronic health records from a tertiary medical centre in Taiwan (2020–2025). A total of 6,089 patients undergoing TKA (n = 4,311) or THA (n = 1,778) were included. Twenty-eight preoperative features were extracted, including haematological indices, demographic variables, comorbidities, and medication use. Four machine learning models (Logistic Regression, Random Forest, XGBoost, LightGBM) were trained using 5-fold stratified cross-validation on an 80% development set and evaluated on a 20% holdout set. Model calibration was assessed using the Hosmer-Lemeshow test. SHAP values were used to interpret feature contributions globally and by surgical subgroup. Results: The overall transfusion rate was 18.7% (TKA: 15.3%; THA: 27.1%). LightGBM achieved the highest holdout AUROC (0.751; 95% CI: 0.714–0.787), with good calibration (Hosmer-Lemeshow p = 0.297), Brier Score of 0.128, and NPV of 0.891 at the Youden Index threshold. SHAP analysis identified preoperative haemoglobin (Hb), surgery type, and tranexamic acid (TXA) use as the top predictors. A non-linear relationship was observed: transfusion risk increased sharply when preop Hb fell below 12 g/dL. Conclusions: An interpretable LightGBM model demonstrated reliable discrimination and calibration for perioperative transfusion prediction. The high NPV suggests this model can safely identify low-risk patients, potentially reducing unnecessary blood preparation in orthopaedic surgical settings. Key messages: • A LightGBM model predicts perioperative transfusion risk in TKA/THA with an AUROC of 0.751 and NPV of 0.891, enabling safer reduction of unnecessary blood preparation. • Preoperative haemoglobin, surgery type, and TXA use are the strongest predictors of transfusion; risk rises sharply when Hb falls below 12 g/dL.
SN - 1101-1262
SN - 1464-360X
ER -