Abstract: TH-PO1116
Predicting Deceased Donor Kidney Transplant Outcomes: Comparing KDRI/KDPI with Machine Learning
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
Author
- Pahl, Eric, OmniLife, Coralville, Iowa, United States
Background
Kidney transplantation is an effective cure for patients suffering from end-stage renal disease. Kidney transplantation is cost-effective, provides a significant survival benefit, and improves the quality of life for patients. One limitation on kidney transplantation is the appropriate assessment of donor quality, for which several indices have been created.
Methods
Machine learning methods (MLM) were compared to kidney donor risk index (KDRI aka KDPI) for the ability to predict graft failure by 12, 24, and 36 months after deceased donor kidney transplantation (DDKT). The MLM model, an ensemble of thousands of randomly generated decision trees, was trained with the same data initially used to develop KDRI.
Results
An MLM trained with the readily available recipient and donor variables performs significantly better than KDRI/KDPI when predicting graft failure by 12, 24, and 36 months after DDKT. When comparing equal prediction failure rates of 10%, MLM successfully predicted 126% more successful DDKTs (an additional 2,148) than KDRI/KDPI from 1995-2005. Over the entire ROC curve, the MLM performed statistically significantly better c-statistic than KDRI/KDPI in all predictions.
Conclusion
Using MLM, many high-KDRI kidney offers resulted in thousands of successful patient outcomes without increasing risk of predicted graft failure. The MLM provided a significant improvement over KDRI for the assessment of kidney offers and give clinical professionals an improved basis for making the critical decisions. This work lays the foundation for future MLM in organ transplantation and describes the steps to measure, analyze, and validate future models.