Characterization of an integrative prognostic score for US patients taken from the DART study
Jonathan S. Bromberg1, Roy Bloom2, Puneet Sood3.
1University of Maryland, Baltimore, MD, United States; 2Penn Medicine, Philadelphia, PA, United States; 3University of Pittsburgh Medical Center, Pittsburgh, PA, United States
Background: iBox is a validated cloud-based software as a service (SaaS) algorithm that provides a predictive analysis of the post-transplant patient, quantifying the risk of kidney loss based upon multiple pre-determined clinical factors in kidney transplant scenarios at any time point following surgery as long as the inputs necessary are available.
Mandatory inputs include: 1. time from transplant to risk evaluation; 2. estimated GFR (mL/min/1,73m²); and 3. proteinuria (g/g of creatinine/total protein). Additional inputs that improve the accuracy include DSA MFI Banff scores (g,i,t,v,ptc,cg,ci,ctcv,ah,C4d) or other diagnoses based on pathology (e.g., AKI, BKVN, TCMR, borderline lesions, ABMR, recurrence). The output result includes allograft loss probabilities for 1, 3 and 5 years from evaluation time, and potential stratification based on renal function trajectory prediction.
The objective of this study was to characterize the iBox algorithm using a smaller cohort of patients from 14 centers who were prospectively surveyed in the DART study over 12 months (ClinicalTrials.gov Identifier: NCT02424227) to assess whether the predictive ability of iBox matched the outcome. Donor derived cell free DNA (dd-cfDNA), not currently included in the iBox algorithm, was independently considered as a predictor of outcome at 12 months.
Methods: 185 patients from DART were identified as having sufficient data to complete the mandatory inputs required for iBox. Within 12 months of follow up, these patients had a total of 9 acute rejection episodes included in the analysis. iBox scores were calculated at the 7 time points used in the surveillance protocol: months 1,2,3,4,6,9 and 12. This population was representative of patients followed through standard of care practice in US kidney transplant programs.
Results: Inclusion of the most comprehensive set of model parameters increasing from the 3 mandatory inputs boosts statistical power for the prognostication algorithm. DART data did not have complete parameters for iBox at each time-point; for example biopsies were not done at each point. Of the patient data set used, it was 19.5% complete for the fully loaded iBox model. Using the best fit model a C-statistic of 0.83 was obtained with a variance of 19%. iBox 1-year prognostication did not correlated with dd-cfDNA (correlation=-0.1213, p = 0.083).
Conclusion: As transplant moves toward increasingly relying on predictions made by computer models to justify clinical decisions, knowing the iBox algorithm performs as expected in smaller cohorts when considering diagnosis and 1-year prognosis, suggests clear translational value. The inclusion of dd-cfDNA considering allograft injury may be a valuable addition to evolve this algorithm as we improve understanding as well as consider the impact of medical interventions and response to treatments on graft failure risk.
There are no comments yet...