Adaptation of Random Survival Forests for Predicting Interval Censored Outcome Using a Longitudinal Biomarker: Application to Tacrolimus and Antibody Formation in Kidney Transplant

• Activity Number: 522 - Contributed Poster Presentations: Biometrics Section
• Type: Contributed
• Date/Time: Wednesday, July 31, 2019 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Kaci Pickett* and Krithika Suresh and Kristen Campbell and Elizabeth Juarez-Colunga
• Companies: and University of Colorado and University of Colorado and University of Colorado Denver
• Keywords: Random Survival Forest; Interval Censored Time-to-Event outcome; Predictive Model; Longitudinal

Abstract

The immunosuppressant drug Tacrolimus (TAC) is given post kidney transplant to prevent de novo donor specific antibodies (dnDSA), which are thought as an early warning sign of kidney allograft rejection. It is hypothesized that TAC levels over time, in addition to numerous baseline covariates, are predictive of dnDSA. Using data from the University of Colorado Transplant Center, we build a machine learning approach to predict dnDSA using random forests. To employ the traditional machine learning approaches for survival data we first address (i) the interval censored nature of dnDSA and (ii) the longitudinal aspect of TAC. Imputation methods for exact time-to-dnDSA within the censoring interval include using the midpoint of the interval and imputing based on the nonparametric estimator of the survival curve. To address (ii) various RSF models are considered by fixing a time point t post-transplant, from which the prediction will be developed. We evaluate functions of TAC prior to t that are of particular clinical interest as predictors of dnDSA including the median TAC and the coefficient of variation of TAC during the 1 month preceding t.