Evaluating personalized medicine in multi-marker multi-treatment clinical trials: Accounting for heterogeneity

The assessment of the added value when matching the right treatment to the right population based on a molecular profile raises numerous statistical issues. Due to the low prevalence of potential molecular predictive factors of response to treatment as well as of the existence of many types of histology in oncology, it is often impossible to carry out a separate trial for each histology and molecular profile combination. Instead, several contemporary randomized clinical trials investigate the efficacy of algorithms that combine multiple treatments with multiple molecular markers. Some of them focus on a single histology, whereas other are histology-agnostic and test whether selecting the treatment based on biology is superior to selecting the treatment based on histology. Several important sources of variability are induced by these types of trials. When this variability also concerns the treatment effect, the statistical properties of the design may be strongly compromised. In this chapter, using the randomized SHIVA trial evaluating personalized medicine in patients with advanced cancers as example, we present strengths and pitfalls of designs and various analysis tools. In particular, we illustrate the lack of power in the case of an algorithm being partially erroneous, the necessity to use randomized trials compared to designs where the patient is his or (her) own control, and propose a modeling approach to account for heterogeneity in treatment effects at the analysis step.