The use of a predictive statistical model to make a virtual control arm for a clinical trial

Abstract

Background Randomized clinical trials compare participants receiving an experimental intervention to participants receiving standard of care (SOC). If one could predict the outcome for participants receiving SOC, a trial could be designed where all participants received the experimental intervention, with the observed outcome of the experimental group compared to the prediction for those individuals. Methods We used the CancerMath calculator to predict outcomes for participants in two large clinical trials of adjuvant chemotherapy for breast cancer: NSABPB15 and CALGB9344. NSABPB15 was the training set, and we used the modified algorithm to predict outcomes for two groups from CALGB9344: one which received standard of care (SOC) chemotherapy and one which received paclitaxel in addition. We made a prediction for each individual CALGB9344 participant, assuming each received only SOC. Results The predicted outcome for the group which received only SOC matched what was observed in the CALGB9344 trial. In contrast, the predicted outcome for the group also receiving paclitaxel was significantly worse than what was observed for this group. This matches the conclusion of CALGB9344 that adding paclitaxel to SOC improves survival. Conclusion This project proves that a statistical model can predict the outcome of clinical trial participants treated with SOC. In some circumstances, a predictive model could be used instead of a control arm, allowing all participants to receive experimental treatment. Predictive models for cancer and other diseases could be constructed using the vast amount of outcomes data available to the federal government, and made available for public use.

Conditional distribution modeling as an alternative method for covariates simulation: comparison with joint multivariate normal and bootstrap techniques

“…This suggests that study-specific nuisance covariates, such as HS vs. patients, should be handled carefully before embarking in a CTS effort. [24], the use of predictive statistical models to simulate virtual responses to standard-of-care treatments has recently been proposed as a way to generate synthetic controls [25]. In the case of rare diseases and small populations in general (e.g.…”

“…Synthetic controls are formally defined as statistical methods that can be used to evaluate the comparative effectiveness of an intervention using external control data [23]. Although synthetic controls are usually based on collected real evidence [24], the use of predictive statistical models to simulate virtual responses to standard-of-care treatments has recently been proposed as a way to generate synthetic controls [25]. In the case of rare diseases and small populations in general (e.g.…”

Virtual controls as an alternative to randomized controlled trials for assessing efficacy of interventions

“…Switchenko et al [ 48 ] checked the method of virtual controls using data from a trial of paclitaxel as adjuvant chemotherapy for breast cancer. A randomized trial [ 49 ] had tested paclitaxel plus the standard of care versus the standard of care alone.…”

Conditional distribution modeling as an alternative method for covariates simulation: Comparison with joint multivariate normal and bootstrap techniques

“…23 Although synthetic controls are usually based on collected real evidence, 24 the use of predictive statistical models to simulate virtual responses to standard-of-care treatments has recently been proposed as a way to generate synthetic controls. 25 In the case of rare diseases and small populations in general (e.g., pediatrics), where the use of control groups is often hampered by feasibility and ethical hurdles, simulated synthetic controls represent a promising alternative, as also indicated by their increased regulatory acceptance. 26 To conclude, the present analysis revealed that if CTS is used to simulate within the range of the observed distribution-for example, when studies in the target population are already available from other sources-BS can be considered as the preferred method for covariate simulation, particularly because it is able to guarantee the physiological plausibility of the simulated covariates.…”