Short-Term Outcome Prediction Model in Deceased Donor Kidney Transplant Recipients
1OSU, Columbus
2UVA, Charlottesville
3Montefiore, New York
4NWU, Chicago.
Meeting: 2018 American Transplant Congress
Abstract number: B101
Keywords: Donors, Genomic markers, Kidney transplantation, marginal, Outcome
Session Information
Session Name: Poster Session B: Kidney Deceased Donor Allocation
Session Type: Poster Session
Date: Sunday, June 3, 2018
Session Time: 6:00pm-7:00pm
Presentation Time: 6:00pm-7:00pm
Location: Hall 4EF
Background: Lack of accurate outcome prediction models may be a reason for the increase in organ discard rate that further burdens the growing kidney transplant (KT) wait list. Herein, we applied machine learning methods to identify molecular features from deceased donor (DD) pre-implant (PI) biopsies for more accurately predicting short-term KT outcome. This is a first step towards developing a composite scoring system.
Methods: Gene expression was done in 189 PI biopsies from unique KT donors and the significance of association and prediction accuracy of KDRI with short-term outcome, defined as glomerular filtration rate (eGFR) at one month categorized as high (>40) vs. low (<=40), was estimated. A penalized logistic regression model was fitted for predicting high versus low eGFR with KDRI included in the model. Machine learning was then incorporated in data analysis to develop a model with improved predictive performance.
Results: KDRI was significantly associated with short-term outcome (P=0.029) in the 189 DD KT recipients. The area under ROC curve was 0.626 for prediction of short-term outcome. However, when fitting a penalized logistic regression model predicting high vs low eGFR that included KDRI in the model, the addition of expression levels from PI biopsies for 12 probe sets yielded an AUC of 0.811 (Sensitivity 0.74; Specificity 0.75). The issue of severe imbalance between the two classes to be predicted (50 (26.2%) subjects with low eGFR vs 139 (72.8%) subjects with high GFR at one month post-KT) that plagues the development of more accurate prediction model was overcome by employing oversampling of the minority class combined with the machine learning method random forests. We determined the optimal tuning parameter for our random forest by performing 10-fold cross-validation. Then, we fit our random forest model to predict high vs. low GFR at one month. Using observations not included in the fitting procedure (out-of-bag observations), our misclassification rate was only 6.12% (Sensitivity 0.935; Specificity 0.942). Interestingly, among the 19380 possible predictors, KDRI ranked 19239th with respect to variable importance defined as the mean decrease in accuracy.
Conclusions: A panel of PI molecular markers were identified that together with clinical parameters predict short-term outcomes more accurately than scoring systems currently in place.
CITATION INFORMATION: Archer K., Zhang Y., Bontha S., Akalin E., Gallon L., Maluf D., Mas V. Short-Term Outcome Prediction Model in Deceased Donor Kidney Transplant Recipients Am J Transplant. 2017;17 (suppl 3).
To cite this abstract in AMA style:
Archer K, Zhang Y, Bontha S, Akalin E, Gallon L, Maluf D, Mas V. Short-Term Outcome Prediction Model in Deceased Donor Kidney Transplant Recipients [abstract]. https://atcmeetingabstracts.com/abstract/short-term-outcome-prediction-model-in-deceased-donor-kidney-transplant-recipients/. Accessed October 9, 2024.« Back to 2018 American Transplant Congress