Abstract: TH-PO1118
Optimization of Machine Learning Models for Predicting Delayed Graft Function in Renal Allografts
Session Information
- Transplantation: Clinical - Predictors of Outcomes - Biomarkers and Beyond
November 07, 2019 | Location: Exhibit Hall, Walter E. Washington Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Transplantation
- 1902 Transplantation: Clinical
Authors
- Jen, Kuang-Yu, University of California, Davis, Sacramento, California, United States
- Yen, Felicia, University of California, Davis, Sacramento, California, United States
- Sageshima, Junichiro, University of California, Davis, Sacramento, California, United States
- Rashidi, Hooman H., UC Davis School of Medicine, Sacramento, California, United States
Background
Delayed graft function (DGF) is associated with worse short- and long-term renal allograft outcomes. Several groups have previously developed models that compute the theoretical risk of DGF for allograft recipients using standard statistical inference methods. In this study, we apply automated computational algorithms to generate tens of thousands of DGF prediction machine learning models based on donor characteristics alone. In this en masse approach, we are able to empirically optimize these machine learning models for the prediction of DGF.
Methods
Deceased donor data available from UNOS for 1,694 renal transplants at our center from 2010-2018 were used in this study, which included various elements of demographics, medical history, and circumstances of death. Cold ischemia time (CIT) and KDPI were included as well. The number of cases was further trimmed randomly to achieve a 50%/50% split in DGF-positive and negative cases, with a final total of 922 cases. These data were used to create 10 runs for each specific parameter combination [each using 90% (n=830) of cases for training phase and 10% (n=92) of cases for each run’s validation test] to generate a total of 45,980 unique models within these parameter combinations on 4 distinct machine learning algorithms (logistic regression, k-nearest neighbor, support vector machine, random forest). Models were also produced with fewer donor features, KDPI alone, CIT alone, and KDPI with CIT. The mean accuracy, standard deviation, and area under the curve (AUC) for the best models were calculated.
Results
Of the 45.980 models generated, the best performing models had an accuracy of 74% (5.7) and AUC of 77. A common theme to these optimized models was that they excluded KDPI as a feature but included CIT. KDPI alone performed poorly (accuracy 49-57%; AUC 0.51-0.55). CIT alone was also suboptimal (accuracy 57-63%; AUC 0.56-0.62).
Conclusion
Machine learning algorithms can help to produce improved and optimized prediction models for DGF.