Integrating Clinical and Histopathological Data to Predict Delayed Graft Function in Kidney Transplant Recipients Using Machine Learning Techniques
Issued Date
2024-12-01
Resource Type
eISSN
20770383
Scopus ID
2-s2.0-85213290030
Journal Title
Journal of Clinical Medicine
Volume
13
Issue
24
Rights Holder(s)
SCOPUS
Bibliographic Citation
Journal of Clinical Medicine Vol.13 No.24 (2024)
Suggested Citation
Tirasattayapitak S., Ratanatharathorn C., Thotsiri S., Sutharattanapong N., Wiwattanathum P., Arpornsujaritkun N., Sirisopana K., Worawichawong S., Rostaing L., Kantachuvesiri S. Integrating Clinical and Histopathological Data to Predict Delayed Graft Function in Kidney Transplant Recipients Using Machine Learning Techniques. Journal of Clinical Medicine Vol.13 No.24 (2024). doi:10.3390/jcm13247502 Retrieved from: https://repository.li.mahidol.ac.th/handle/20.500.14594/102597
Title
Integrating Clinical and Histopathological Data to Predict Delayed Graft Function in Kidney Transplant Recipients Using Machine Learning Techniques
Corresponding Author(s)
Other Contributor(s)
Abstract
Background: Given the significant impact of delayed graft function (DGF) on transplant outcomes, the aim of this study was to develop and validate machine learning (ML) models capable of predicting the risk of DGF in deceased-donor kidney transplantation (DDKT). Methods: This retrospective cohort study was conducted using clinical and histopathological data collected between 2018 and 2022 at Ramathibodi Hospital from DDKT donors, recipients, and post-implantation time-zero kidney biopsy samples to develop predictive models. The performance of three ML models (neural network, random forest, and extreme gradient boosting [XGBoost]) and traditional logistic regression on an independent test data set was evaluated using the area under the receiver operating characteristic curve (AUROC) and Brier score calibration. Results: Among 354 DDKT recipients, 64 (18.1%) experienced DGF. The key contributing factors included a donor body mass index > 23 kg/m2, donor diabetes mellitus, a prolonged cold ischemia time, a male recipient, and an interstitial fibrosis/tubular atrophy score of 2–3 in the time-zero kidney biopsy sample. The random forest model had a specificity of 99.96% and an AUROC of 0.9323, the neural network model had a specificity of 97.43% and an AUROC of 0.844, and the XGBoost model had a specificity of 99.81% and an AUROC of 0.989. A traditional statistical model had a specificity of 84.4% and an AUROC of 0.769. Conclusions: Predictive models, especially XGBoost models, have potential as tools for assessing DGF risk post-DDKT, guiding acceptance decisions, and avoiding risky biopsy, and they may be crucial in resource-limited settings.