Artificial Intelligence-Based Prediction of Postoperative Blood Loss in Cardiac Surgery in Samutsakhon Hospital

Abstract

Objective: To develop and compare the performance of artificial intelligence–based models for predicting blood loss volume within the first 4 hours after open heart surgery and for identifying patients at risk of postoperative bleeding greater than 400 mL, and to evaluate their potential application as a clinical decision-support tool

Methods: A retrospective descriptive study was conducted in 530 patients who underwent open heart surgery at Samutsakhon Hospital between January 1, 2019, and April 30, 2025. Preoperative, intraoperative, and postoperative clinical data were used to develop supervised machine learning models, including Random Forest, HistGradientBoosting, and XGBoost. The models were designed to predict both postoperative bleeding greater than 400 mL within 4 hours (classification task) and continuous blood loss volume (regression task). Model development was performed using Python programming language. Hyperparameter tuning was conducted using 3-fold cross-validation, and model performance was subsequently evaluated using 5-fold cross-validation. Classification performance was assessed using precision, recall, F1-score, accuracy, and area under the receiver operating characteristic curve (ROC AUC), whereas regression performance was evaluated using mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²).

Results: Patients with postoperative bleeding greater than 400 mL within 4 hours had lower preoperative platelet counts, longer aortic cross-clamp time, longer cardiopulmonary bypass time, and significantly different surgical procedures (χ² = 9.43; p = .02), with combined coronary artery bypass grafting (CABG) and valve surgery occurring more frequently in the high-bleeding group. Among the evaluated models, XGBoost demonstrated the best classification performance, achieving a precision of 0.783, recall of 0.684, F1-score of 0.728, accuracy of 0.909, and ROC AUC of 0.923. For continuous blood loss prediction, XGBoost achieved a MAE of 85.17 mL, RMSE of 119.80 mL, and R² of 0.620. Predicted values showed good agreement with observed blood loss, although a tendency toward underestimation was observed in patients with extremely high bleeding volumes.

Conclusion: The XGBoost model demonstrated strong potential as a clinical decision-support tool for assessing the risk and volume of postoperative bleeding following open heart surgery. Its application may facilitate perioperative blood management and improve clinical decision-making regarding blood preparation and resource utilization.