Use of predictive models in CNS diseases

The role of placebo response in the efficacy outcome assessment in viloxazine extended-release pivotal trials in paediatric subjects with attention-deficit/hyperactivity disorder

AIMS

Four Phase 3 studies evaluated efficacy and safety of viloxazine extended-release in the treatment of attention-deficit/hyperactivity disorder (ADHD). The primary efficacy objective-change from baseline in ADHD Rating Scale-5 (ADHD-RS-5) Total score at end of study (EOS)-was not met in one of the studies (812P304). A band-pass analysis was performed to evaluate the impact of placebo response on the results.

METHODS

The distribution of placebo response at EOS of each trial was evaluated. The 2.5^th and 97.5^th percentiles of the distribution of ADHD-RS-5 Total score were used as boundaries for the band-pass analysis. An independent mixed model for repeated measures analysis was conducted for each trial using all eligible data (active and placebo) from the total and band-pass filtered populations.

RESULTS

The 2.5^th and 97.5^th percentiles at EOS were 3.5 and 53.5, respectively. Application of the band-pass filter (filtering out all subjects [active, n = 305 (32.1%) and placebo, n = 134 (33.5%)] of clinical sites with placebo scores <3.5 or >53.5) revealed statistically significant improvement at the primary endpoint (600-mg/d viloxazine ER vs. placebo) in Study 812P304 (mean [confidence interval] = 4.9537 [0.5405-9.3669]), previously masked by a high placebo response (mean [confidence interval] = 3.5756 [-0.3332-7.4844]). The outcome of the analysis indicated that the impact of the band-pass adjustment is greater when placebo response is higher.

CONCLUSION

This analysis indicated that a higher placebo response in Study 812P304 confounded the assessment of treatment effect. Application of the band-pass methodology confirmed the positive results of the 3 prior studies and the signal detection confounder in the fourth study.

Model-Informed Approach to Assess the Treatment Effect Conditional to the Level of Placebo Response

One of the most important reasons for failure of placebo-controlled randomized controlled clinical trials (RCTs) is the lack of appropriate methodologies for detecting treatment effect (TE; difference between placebo and active treatment response) in the presence of excessively low/high levels of placebo response. Although, the higher the level of placebo response in a trial, the lower the apparent detectable TE. TE is usually estimated in a conventional analysis of an RCT as an "apparent" TE value conditional to the level of placebo response in that RCT. A model-informed methodology is proposed to establish a relationship between level of placebo response and TE. This relationship is used to estimate the "typical" TE associated with a "typical" level of placebo response, irrespective of the level of placebo response observed. The approach can be valuable for providing a reliable estimate of TE, for conducting risk/benefit analysis, and for determining dosage recommendations.

A Novel Methodology to Estimate the Treatment Effect in Presence of Highly Variable Placebo Response

One of the main reasons for the inefficiency of multicenter randomized clinical trials (RCTs) in depression is the excessively high level of placebo response. The aim of this work was to propose a novel methodology to analyze RCTs based on the assumption that centers with high placebo response are less informative than the other centers for estimating the 'true' treatment effect (TE). A linear mixed-effect modeling approach for repeated measures (MMRM) was used as a reference approach. The new method for estimating TE was based on a nonlinear longitudinal modeling of clinical scores (NLMMRM). NLMMRM estimates TE by associating a weighting factor to the data collected in each center. The weight was defined by the posterior probability of detecting a clinically relevant difference between active treatment and placebo at that center. Data from five RCTs in depression were used to compare the performance of MMRM with NLMMRM. The results of the analyses showed an average improvement of ~15% in the TE estimated with NLMMRM when the center effect was included in the analyses. Opposite results were observed with MMRM: TE estimate was reduced by ~4% when the center effect was considered as covariate in the analysis. The novel NLMMRM approach provides a tool for controlling the confounding effect of high placebo response, to increase signal detection and to provide a more reliable estimate of the 'true' TE by controlling false negative results associated with excessively high placebo response.

Use of band-pass filter analysis to evaluate outcomes in an antidepressant trial for treatment resistant patients

Band-pass filtering is a novel statistical methodology that proposes that filtering out data from trial sites generating non-plausible high or low levels of placebo response can yield a more accurate effect size and greater separation of active drug (when efficacious) from placebo. We applied band-pass filters to re-analyze data from a negative antidepressant trial (NCT00739908) evaluating CX157 (a reversible and selective monoamine oxidase inhibitor-A) versus placebo. 360 patients from 29 trial sites were randomized to either CX157 treatment (n=182) or placebo (n=178). We applied two filters of<3 or>7 points (filter #1) or<3 and>9 points (filter #2) mean change of the total MADRS placebo scores for each site. Trial sites that had mean placebo MADRS score changes exceeding the boundaries of these band-pass filter thresholds were considered non-informative and all of the data from these sites were excluded from the post-hoc re-analysis. The two band-pass filters reduced the sample of informative patients from 353 patients in the mITT population to 62 in filter #1 and 152 in the filter #2 group. The placebo response was reduced from 31.1% in the mITT population to 9.4% with filter #1 and 20.8% with filter #2. MMRM analysis revealed a non-statistically significant trend of p=0.13 and 0.16 for the two filters in contrast to the mITT population (p= 0.58). Our findings support the band-pass filter hypothesis and highlight issues related to site-based scoring variability and inappropriate subject selection that may contribute to trial failure.