PS38: Missing Data in Bridging Studies of Companion Diagnostics

A companion diagnostic (CDx) is a medical device (test) that provides information that is essential for the safe and effective use of a corresponding drug or biological product. To evaluate the clinical benefit of a CDx test, one may stratify the treatment effect on an efficacy endpoint by the test result. The analysis population for stratified treatment effects should be the intent-to-treat (ITT)/intent-to-diagnose (ITD) population. Commonly, participants are enrolled into the trial using clinical trial assays (CTA), or local tests, that are not the market ready CDx candidate, with samples saved to be retrospectively tested by the candidate, and the test results to be used to bridge from CTA to CDx. Unfortunately, missing CDx test results often arise: • The very common trial design for CDx validation is the biomarker-targeted design where only CTA+ patients are eligible for enrollment and treatment. In this design, efficacy data are missing for the (CDx+, CTA-) population because they aren't enrolled. • Some of the samples saved for CDx re-testing lack sufficient specimen or relevant tissue content in good quality to obtain a reliable result from the CDx test. For some of the samples saved for CDx re-testing, the CDx test result is invalid. Because CDx results may be missing not at random in the clinical trial, which may introduce bias in terms of interpreting the results from the observed data or lead the results lack of generalizability to the intent-to-diagnose population, additional steps are recommended to be taken to address the missingness, such as: • To impute the missing efficacy data of the (CDx+, CTA-) population based on hypothetical assumptions for a sensitivity analysis, or from an empirical CDx-efficacy relationship found in literature, a past study or the existing data. • To impute the CDx test results from the likely correlated data structure, or known causes. Due to the CDx test results missing, the observable treatment effect conditional on the available CDx test results is estimated from a subset to the ITT population, which may need to be adjusted by the other covariates to the treatment effect. Through the process, the potential relationship found between the CDx test results and the other covariates may facilitate imputing the CDx test results. • To design and conduct an external concordance study between the CDx candidate and CTA (or an approved CDx). In the external concordance study, the population is to be representative to the intent-to-diagnose population, and it's to demonstrate the interchangeability of the two tests and the similarity between their test performances. In the era of artificial intelligence/machine learning (AI/ML), although with challenges in robustness, interpretability and reasoning, some AI/ML models may be useful tools in imputing missing data, selecting or monitoring trial populations. This session will provide perspectives from FDA, pharmaceutical and medical device industries.