Sequential likelihood ratios and e-processes in the analysis of the RENOVATE trial

Adaptive platform trials enrolling heterogeneous populations face a critical challenge: when treatment effects differ across subgroups and enrollment is non-concurrent, pooled analyses can produce misleading results due to shifting patient composition over time.

To quantify treatment effect heterogeneity in the RENOVATE trial using evidential methods and e-processes, and to demonstrate how temporal shifts in enrollment composition can create statistical artifacts in sequential monitoring.

Secondary analysis of the RENOVATE trial, which randomized 1766 adults with acute respiratory failure to high-flow nasal oxygen (HFNO) versus noninvasive ventilation (NIV) across five populations: non-immunocompromised hypoxemia (n = 485), immunocompromised hypoxemia (n = 50), COPD exacerbation (n = 77), cardiogenic pulmonary edema (n = 272), and COVID-19 (n = 882). We computed sequential likelihood ratio (SLR) processes within each group for the primary outcome (death or intubation at 7 days), testing a 5% absolute risk reduction hypothesis. We compared group-specific trajectories with pooled analysis to visualize how enrollment composition influenced evidence accumulation. Sensitivity analyses used conditional e-processes (which eliminate the baseline rate parameter) and randomization-based e-processes (assumption-free).

Treatment effects varied substantially across populations. Cardiogenic edema showed strong evidence of HFNO benefit (absolute risk difference -11.0%; S-3 interval -15.8% to -3.8%; final support S = 3.36). COVID-19 showed a point estimate suggesting harm (+4.3%; S-3 interval -1.3% to + 9.8%; S = -3.23). The pooled SLR displayed a V-shaped artifact, with support dropping to S = -5.5 during the COVID-dominated enrollment period, then reversing to S = +5.6 as lower-risk patients entered. Alternative analyses were aligned with SLR but did not suffer from artifact interpretation due to baseline risk change.

Sequential evidential analysis reveals substantial treatment effect heterogeneity, which is masked by pooled analysis, with enrollment composition mechanistically driving the evidence trajectories. E-processes provide a diagnostic tool for platform trials, making visible the interaction between enrollment dynamics and treatment heterogeneity that conventional pooled estimates cannot reveal.

NCT03643939.

Brazilian Ministry of Health.