Multicenter machine learning-based construction of a model for predicting potential organ donors and validation with decision curve analysis
10.12464/j.issn.1674-7445.2025208
- VernacularTitle:多中心机器学习构建预测潜在器官捐献者的模型与决策曲线验证研究
- Author:
Xu WANG
1
;
Wenxiu LI
1
;
Fenghua WANG
1
;
Shuli WU
1
;
Dong JIA
1
;
Xin GE
1
;
Zhihua SHAN
1
;
Tongzuo LI
1
Author Information
1. Office of Organ Donation, Department of Medical Affairs, the First Affiliated Hospital of Harbin Medical University, Harbin 150001, China.
- Publication Type:OriginalArticle
- Keywords:
Multicenter machine learning;
Potential organ donor;
Predictive model;
Decision curve analysis;
Extreme gradient boosting;
Organ donation assessment;
Logistic regression model;
Random forest model
- From:
Organ Transplantation
2026;17(1):106-115
- CountryChina
- Language:Chinese
-
Abstract:
Objective To evaluate the predictive value of different machine learning models constructed in a multicenter environment for potential organ donors and verify their clinical application feasibility. Methods The study included 2 000 inpatients admitted to five domestic tertiary hospitals from January 2020 to December 2023, who met the criteria for potential organ donation assessment. They were randomly divided into a training set and an internal validation set (7∶3). Another 300 similar patients admitted to the First Affiliated Hospital of Harbin Medical University from January 2024 to April 2025 were included as an external validation set. The area under the curve (AUC), sensitivity, specificity, accuracy and F1-score of three models were compared, and the consistency of the potential organ donor determination process was tested. Multivariate logistic regression analysis was used to identify predictive factors of potential organ donors. Decision curve analysis (DCA) was employed to verify the resource efficiency of each model, and the threshold interval and intervention balance point were assessed. Results Apart from age, there were no significant differences in other basic characteristics among the centers (all P>0.05). The consistency of the potential organ donor determination process among researchers in each center was good [all 95% confidence interval (CI) lower limits >0]. In the internal validation set, the XGBoost model had the best predictive performance (AUC=0.92, 95% CI 0.89-0.94) and the best calibration (P=0.441, Brier score 0.099). In the external validation set, the XGBoost model also had the best predictive performance (AUC=0.91, 95% CI 0.88-0.94), outperforming logistic regression and random forest models. Multivariate logistic regression showed that mechanical ventilation had the greatest impact (odds ratio=2.06, 95% CI 1.54-2.76, P<0.001). DCA indicated that the XGBoost model had the highest net benefit in the threshold interval of 0.2-0.6. The “treat all” strategy only had a slight advantage at extremely low thresholds. The recommended threshold interval, which balances intervention costs and clinical benefits, considers ≥50% positive predictive value (PPV) and ≤50 referrals per 100 high-risk patients. Conclusions The XGBoost model established in a multicenter environment is accurate and well-calibrated in predicting potential organ donors. Combined with DCA, it may effectively guide the timing of clinical interventions and resource allocation, providing new ideas for the assessment and management of organ donation after brain death.
