Construction and validation of a perioperative blood transfusion model for patients undergoing total hip arthroplasty with osteonecrosis of the femoral head based on machine learning

Abstract

Background: This study aimed to construct a predictive model utilizing multiple machine learning (ML) models to estimate the likelihood of perioperative blood transfusion in patients with osteonecrosis of the femoral head (ONFH) who underwent total hip arthroplasty (THA).

Methods: Patients diagnosed with ONFH who underwent THA at our institution between October 2018 and October 2023 were included in the study. Feature selection was conducted using Lasso regression and correlation analysis. An unbiased evaluation framework incorporating nested resampling was established to assess four ML models. A nomogram was subsequently developed based on the selected features.

Results: Seven features were identified, namely blood loss, hemoglobin (HGB) levels, weight, body temperature, systolic pressure, and direct bilirubin. Four ML models were constructed based on these features. The area under the curve (AUC) values for Random Forest, Extreme Gradient Boosting, Light Gradient Boosting Machine, and Logistic Regression (LR) were 1.00, 1.00, 1.00, and 0.93 in the internal validation set, and 0.89, 0.90, 0.88, and 0.91 in the external test set, respectively. Furthermore, a nomogram model based on LR was developed using the aforementioned seven features, yielding AUC values of 0.95 and 0.90 for the training and test sets, respectively, thereby surpassing the AUC values of preoperative HGB levels (0.80 and 0.76).

Conclusion: Both the ML models and the nomogram exhibit significant potential for forecasting the likelihood of perioperative blood transfusion in patients with ONFH undergoing THA, which may aid clinicians in improving the accuracy of blood transfusion predictions.

Keywords: blood transfusion; machine learning; nomogram model; osteonecrosis of the femoral head; risk prediction; total hip arthroplasty.

Figure 1

Lasso regression and Spearman's correlation analyses. (A) Distribution map illustrating the Lasso coefficients for all variables. (B) Identification of variables through Lasso regression analysis. (C) Correlation analysis conducted among the variables selected by Lasso regression.

Figure 2

ROC curves for four machine learning models evaluated on the internal validation set. (A) The ROC curve for the Random Forest model. (B) The ROC curve for the Extreme Gradient Boosting model. (C) The ROC curve for the Light Gradient Boosting Machine model. (D) The ROC curve for the Logistic Regression model.

Figure 3

Box plot comparing four machine learning models on the external test set. (A) AUC values, (B) ACC values, (C) Recall values, and (D) CE values.

Figure 4

Statistical visualizations of the SHAP analysis. (A) An ordered plot illustrating the importance of variables in the SHAP analysis; (B) SHAP value contribution graph for a single sample's indicators.

Figure 5

Nomogram model construction based seven features. ^*Denotes the magnitude of the P-value.

Figure 6

The receiver-operating characteristic curves, calibration plots, and decision curve analysis for the nomogram model in the training and validation sets are represented by (A–F), respectively.