Pooling trial data to identify heterogeneity and characteristics of patients most likely to respond to treatment: a causal forest approach

Abstract

AbstractBackgroundA potential explanation for the limited success of past therapeutic trials for Alzheimer’s disease (AD) lies in disease heterogeneity. We set out to pool and jointly analyze data from completed trials using a hypothesis‐free machine learning approach, aiming to identify participant characteristics related to increased responsiveness to treatment. Considering methodological difficulties, we also aim to describe the pooling and analytical process.MethodWe requested data from fifteen completed phase I–III trials investigating several compounds in AD individuals and were granted access to six trial datasets using compounds LY450139 (semagacestat), GSK933776, and BI409306 (with donepezil) with 2,397 participants (73.3±7.8 years old, 55% female, Table 1), through the Vivli platform. Causal Forest analyses, a machine learning technique that extracts a minimal subset of patient characteristics that can explain heterogeneity in outcomes, were used with change on (1) ADAS‐Cog and (2) MMSE as outcome, accounting for treatment group (placebo vs. compound).ResultThe included trial datasets showed considerable differences in available variables and data structure, constraining the amount of data that may be used in joint analysis. We were able to pool and analyze twenty‐seven characteristics shared across trial datasets in the model, covering demographics, vital signs, MRI volumetrics, baseline cognitive and functional performance, apolipoprotein ε4‐carriership and drug compound. Causal forest analyses did not reveal heterogeneity in pooled treatment effects on ADAS‐Cog (estimate for differential forest prediction: ‐2.25, p = 0.96) or MMSE (‐2.97, p = 0.98), indicating that these patient characteristics were not predictive for increased responsiveness to treatment. Stratifying by compound, similar results were found.ConclusionWe took an innovative approach and provided a methodological basis to examine heterogeneity in treatment effects by pooling existing trial datasets. Our results imply that responsiveness to treatment may not be well‐explained by the patient characteristics to our disposal, or that different compounds should be investigated. More research into other factors possibly related to treatment response is needed, e.g., fluid biomarkers and genetics. Based on our experience, we recommend that trial sponsors improve collaboration and data access, as we work towards personalized medicine by identifying which therapeutic approach works best for whom.