Creating a synthetic control arm from previous clinical trials: Application to establishing early end points as indicators of overall survival in acute myeloid leukemia (AML).

Abstract

7021 Background: Clinical trials of experimental drugs require controls. Concurrently randomized controls are the gold standard for judging drug effect. Historical controls are not ideal but are much more efficient and economical. Historical controls derived from a single clinical trial have the biases of that trial. Using many trials with comparable end points and eligibility minimizes such bias. Medidata’s archive contains >3000 trials with clinical data rights for deidentified aggregated analyses. We used this resource to develop a synthetic control arm (SCA) for a particular phase I/II single-arm trial in AML. We demonstrate the utility of this approach by addressing a different but equally important issue: establishing early end points as predictors of long term clinical outcomes. Methods: We built an SCA from 7 relapsed/refractory AML trials completed in last 5 yrs. They had similar eligibility criteria as a particular phase I/II trial for an investigational agent. We selected subjects for the SCA who had baseline covariates matching the subjects in the tri.al. Data cleaning and standardization ensured consistency of data fields. The primary outcomes were CR (complete remission) and CRi (CR without hematologic recovery) at 56 days, and overall survival (OS) subsequent to 56 days. Non-CR/non-CRi deaths before 56 days were set to OS=0. We used a landmark analysis to correlate CR and CRi with OS, calculating the hazard ratio (HR) of OS of CR and CRi vs its comparison group. Results: The SCA included 340 subjects (median age 63 yrs, 55% male, 77% White Non-Hispanic, 28% ECOG 0). Results are in this table. Conclusions: The Medidata trial archive is a resource for creating SCAs. The example SCA we created identified well-defined subjects for whom a CR or CRi is associated with longer OS. Investigations of SCAs for other drugs could aid in addressing the types of subjects and drug categories for which CR or CR/CRi predict longer OS. Such information can help build more efficient and more informative adaptive clinical trials. [Table: see text]

Using Electronic Health Records to Derive Control Arms for Early Phase Single‐Arm Lung Cancer Trials: Proof‐of‐Concept in Randomized Controlled Trials

“…[19][20][21][22] Jia et al 22 examined an application in prostate cancer using a modeling approach based on existing trial data with reasonable success. Other studies drawing on aggregate findings from the clinical trial literature have been performed in acute myeloid leukemia 20 and anaplastic lymphoma kinase-targeted Figure 4 Estimates of relative risk across different adjustment methods comparing external control analysis to trial result: randomized controlled trial NCT0236143. To achieve greater balance among baseline covariates for PS analysis methods-"PS IPTW," "PS Cox adjustment," and "PS Stratification"-PS were trimmed by removing nonoverlapping observations from the PS distributions of the experimental and controls arms resulting in slightly modified cohorts in both arms.…”

Estimation of historical control rate for a single arm de-escalation study – Application to the POSITIVE trial

“…Other methods can be explored for evaluating the treatment effect in single-arm trials. The use of synthetic control arm was suggested for a single-arm trial where the data from Medidata’s archive of 340 patients with acute myeloid leukemia were mined and matched with the 16 patients from an experimental arm for analysis [ 11 ]. Patients were matched on ≥4 of 6 baseline criteria at the individual patient level and then combined to estimate the treatment effect, as well as for other exploratory subgroup analyses.…”

Borrowing From Historical Control Data in Cancer Drug Development: A Cautionary Tale and Practical Guidelines

“…The emergence of master protocols in oncology (Renfro and Sargent, 2017) will certainly be of help here, especially in drug development for rare diseases and some cancers. In more general cases, significant recent effort has focused on the selection of "synthetic control" patients from historical populations arising from similar protocols that are matched in some way to the current single-arm intervention patients, where the matching might be determined by the patients' baseline characteristics (Berry et al, 2017). However, even here it can be a strong assumption that there has been no drift in the control rate since the historical studies were conducted, even in the presence of propenity score or other similar statistical adjustment.…”