Abstract: SA-PO1092
Leveraging Machine Learning Methods and Novel Data Sources to Develop Race-Free Algorithms to Predict Deceased Donor Kidney Quality
Session Information
- Transplantation: Clinical - II
November 04, 2023 | Location: Exhibit Hall, Pennsylvania Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Transplantation
- 2102 Transplantation: Clinical
Authors
- Rubin, Jeremy, University of Pennsylvania Perelman School of Medicine, Philadelphia, Pennsylvania, United States
- Zee, Jarcy, University of Pennsylvania Perelman School of Medicine, Philadelphia, Pennsylvania, United States
- Ratcliffe, Sarah J., University of Virginia, Charlottesville, Virginia, United States
- Goldberg, David S., University of Miami School of Medicine, Miami, Florida, United States
- Harhay, Michael O., University of Pennsylvania Perelman School of Medicine, Philadelphia, Pennsylvania, United States
- Parikh, Chirag R., Johns Hopkins Medicine, Baltimore, Maryland, United States
- Vail, Emily Anne, University of Pennsylvania Perelman School of Medicine, Philadelphia, Pennsylvania, United States
- Abt, Peter, University of Pennsylvania Perelman School of Medicine, Philadelphia, Pennsylvania, United States
- Reese, Peter P., University of Pennsylvania Perelman School of Medicine, Philadelphia, Pennsylvania, United States
- Potluri, Vishnu S., University of Pennsylvania Perelman School of Medicine, Philadelphia, Pennsylvania, United States
Background
To help transplant professionals choose appropriate kidneys for patients, the national transplant organization uses the Kidney Donor Risk Index (KDRI) that predicts the risk of graft failure. Unfortunately, the KDRI has modest predictive accuracy (C-statistic 0.6). The KDRI has been criticized for penalizing hepatitis C virus (HCV), given recent therapies, and race, given ethics concerns. We aimed to develop machine learning based, race-free models and longitudinal organ donor lab data to predict 3-year all-cause kidney (graft) survival and delayed graft function (DGF, dialysis in the first week post-transplant).
Methods
Using registry data, we assembled a cohort of adult (≥18 years) recipients of deceased donor kidneys between 5/1/07-12/31/21. We developed models with different combinations of donor and recipient characteristics using standard regression (Cox and logistic regressions) and machine learning algorithms (Random Forest, Ridge, Lasso, Elastic Net).
Results
The final cohort included 162,905 recipients of kidneys from 95,811 donors. Median donor age was 39.5 years (IQR 27-51) and 38% were female. For 3-year kidney survival, the C-statistic of KDRI was 0.59. Removal of donor race and HCV covariates, or inclusion of donor longitudinal data, made little difference, but inclusion of recipient covariates improved the C-statistic to 0.63-0.64 across different algorithms (Figure). For DGF, the KDRI had a C-statistic of 0.60. Refitting models with or without donor race had similar C-statistics of 0.68-0.69. The inclusion of recipient covariates improved the C-statistics for DGF to 0.74-0.75.
Conclusion
These models demonstrated the feasibility of eliminating donor HCV and race without a meaningful decrement in predictive accuracy for kidney survival and DGF outcomes. The addition of recipient characteristics substantially improves prediction of both outcomes.