Explainable machine learning for the assessment of donor grafts in liver transplantation

The shortage of liver grafts compared to recipients necessitates precise organ assessment. This study aimed to develop a Machine learning (ML) model to predict postoperative delayed graft function (DGF) and visualize the decision-making process for clinical application. Data from 5242 donor-recipien...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Hepatology research Jg. 55; H. 6; S. 908 - 921
Hauptverfasser: Zhixing, Liang, Linsen, Ye, Peng, Jiang, Siyi, Dong, Haoyuan, Yu, Kun, Li, Siqi, Li, Yongwei, Hu, Mingshen, Zhang, Wei, Liu, Hua, Li, Shuhong, Yi, Guihua, Chen, Xiao, Xu, Shusen, Zheng, Yang, Yang
Format: Journal Article
Sprache:Englisch
Veröffentlicht: Netherlands Wiley Subscription Services, Inc 01.06.2025
Schlagworte:
Accuracy
Algorithms
Blood levels
Decision making
Learning algorithms
Liver transplantation
Liver transplants
Machine learning
Survival
γ-Glutamyltransferase
eXtreme gradient boosting algorithms
machine learning
SHapley additive exPlaination
delayed graft function
liver transplantation
ISSN:1386-6346, 1872-034X
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:The shortage of liver grafts compared to recipients necessitates precise organ assessment. This study aimed to develop a Machine learning (ML) model to predict postoperative delayed graft function (DGF) and visualize the decision-making process for clinical application. Data from 5242 donor-recipient pairs who underwent liver transplantation (LT) at the top 10 liver transplant centers in China (January 2017 to December 2022) were collected. The dataset was divided into training and validation sets. Sixty-three variables, including demographics, donor characteristics, diagnosis, preoperative lab results, and surgical information were analyzed. The primary outcome was posttransplantation DGF and the second outcome was posttransplantation 1-month and 3-month survival. Recursive feature elimination selected critical variables, and models were built using ML algorithms and logistic regression. Model performance was evaluated by AUC, accuracy, sensitivity, and specificity. The best model was validated with an independent dataset of 394 LT cases (January to June 2023). The SHapley Additive exPlanations package interpreted the top model's decisions. Among 5242 cases, 328 (6.26%) developed DGF, with 15 cases (3.81%) in the external validation set. Thirty critical features were selected. The eXtreme Gradient Boosting algorithm achieved the highest AUC (0.877) and accuracy (0.936) in the internal set, and a comparable AUC (0.776) and accuracy (0.957) in the external set. SHAP analysis identified short perfusion time, high donor serum sodium, excessive bleeding during transplantation, high donor γ-glutamyl transpeptidase, and blood glucose levels as top predictors of post-LT DGF. The proposed model AUC's 1-month survival prediction was 0.841 and the 3-month survival prediction was 0.834. The developed model for predicting postoperative DGF demonstrated excellent predictive performance, aiding clinicians in evaluating donor grafts and making informed decisions.
Bibliographie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
content type line 23
ISSN:1386-6346
1872-034X
DOI:10.1111/hepr.14187