Clinical Trial Data Set Re-use With Statistical Methodologies Tailored for Clinical Trials in Rare Diseases

EPISTOP-IDEAL (EPISTOP clinical trial data set re-use with statistical methodologies tailored for clinical trials in rare diseases).aims to show the usability and capability of the newly developed statistical methodologies for clinical trials in rare diseases and how they can help to better consolidate the results of the randomized clinical trials (RCT). CTs are often performed with standard classical methodologies not specific for rare diseases resulting in a loss of power to show positive effects or in the incapacity to answer all the questions that patients, clinicians and regulatory agencies need to obtain from therapy trials. This call will benefit from an exceptional set of data collected longitudinally in patients with TSC during EPISTOP project (2013-2019) and beyond as the clinical follow-up of patients is continuing following the EPISTOP protocol and using the same drug (vigabatrin) with the same doses range. EPISTOP-IDEAL aims, through a collaboration between EPISTOP coordinator and members and expert statisticians from the IDEAL (Integrated DEsign and AnaLysis of clinical trials in Small population groups, a previous EU project on methodologies developed for trials in small populations), to consolidate the primary end point results of the trial using bias assessment and external data and to strengthen the results of some major secondary end points that were not conclusive because of the small number of the sample. This project is based on collaboration between experts in TSC, a rare disease with high morbidity, and clinical trials methodologies' experts that developed innovative methodologies that can re-evaluate data that might have lacked partly efficiency because it was analysed with classical statistical methodology as RCTs. EPISTOP-IDEAL will aim to re-use the data collected during EPISTOP trial that compared the impact of early (preventive) or late (standard) treatment by Vigabatrin (the drug usually used to treat early onset epilepsies in patients with TSC) on the primary end point of the trial (time to seizure onset) between 2 groups of infants enrolled in this trial before the onset of seizures. In addition to these data, similar data have been acquired since the EPISTOP end of recruitment in 2018. These external data are available in the health institutions involved in this call and will be collected and used to provide additional external data. The EPISTOP-IDEAL project will include 5 partners showing complementarity in expertise and including the coordinator of EPISTOP, the WP leader of the EPISTOP clinical trial, one member of EPISTOP consortium and 2 experts in methodologies that developed the innovative methodologies that will be used in EPISTOP-IDEAL.

Description of the unmet need(s) addressed One hundred and one infants were enrolled in the EPISTOP project. Seven children were later excluded due to misdiagnosis. The intention was to randomize all infants with EEG abnormalities into one of two arms: one received ASM with vigabatrin immediately after epileptic changes on EEG, and the second received the same drug with the same range of dose later, after the onset of seizures. However, in some countries the randomized study in infants was not approved by the ethics committee and the study had to be conducted with two subgroups: observational and randomized subgroups. In both, the same criteria for the diagnosis of epileptic changes on EEG, the same criteria for preventive or standard treatment, and the same ASM (vigabatrin) were used. However, the number of the children participating in the randomized arm of EPISTOP decreased to 53. Moreover, in some children preventive treatment was not feasible due to seizures' early onset before the EEG recording and the randomization and some others dropped-out during 24-month- long follow-up. Therefore, 26 children completed the study in the randomized arm and 24 in the observational subgroup. The primary endpoint of the project was the time to first seizure and even though the groups were smaller than expected, this primary endpoint was met. With respect to the statistical evaluation of the data, the EPISTOP consortium observed various effects in the data, which could not be confirmed yet, because of a gap in standard statistical analysis methodologies. IDEAL's findings are considered to be helpful in this context so that a unique opportunity was to apply to a joint project with the IDEAL's experts. The formed consortium will explore the effect of innovative methodologies in these settings and serves as a nucleus for future studies in Epilepsy as well as similar rare disease areas. Aspects of reanalysis include a) Assess the level of evidence linked to randomization procedures, b) Rigorous use of methods for extrapolation, c) Identify most sensitive response variables and d) Assess and overcome uncertainty in estimate. After EPISTOP enrolment completion, most health institutions involved in EPISTOP continued the use of EPISTOP protocol for the clinical best care of infants with TSC. Therefore, there are patients with TSC followed with the same protocol, using depending on the centre the preventive or the standard treatment and that we can propose as external data for validation of EPISTOP results Description of the proposed innovative methodological analysis

To use the innovative statistical methods to design and analyse small population clinical trials developed under the umbrella of the IDEAL project (Hilgers et al., 2018) the following methods are considered:

1. Development regarding level of evidence linked to randomization procedures b) Rigorous use of methods for extrapolation c) Methods to identify trial most sensitive response variables d) Methods to assess and overcome uncertainty in estimates
2. The data from the randomized clinical trial of the EPISTOP project will be used to investigate the impact of the allocation process on the level of evidence. In particular we will specify a bias model using the biasing policy and relate the model to the allocation sequence resulting from the applied randomization procedure. Based on this we are able to quantify the impact of bias on the study results expressed as p-values as well as confidence intervals (Hilgers et al., 2018). This can be interpreted as uncertainty evaluation of trial results. We will then use the model to derive the biased corrected test result on the primary endpoint (Rückbeil,Hilgers and Heussen, 2019). This is the part of the reanalysis of the randomized trial data. The findings of the reanalysis can be used in two further directions. As we know an estimate of the potential impact of bias on the study result, we will derive recommendations for planning future trials. This concerns the formal evaluation of randomization procedures to identify the recommended procedure protecting against bias in a future study. On the other hand, we use the same approach to quantify the impact of bias in the observational part of the EPISTOP project. If the resulting amount of bias identified in the randomized and the observational trial correspond to each other, we are allowed to pool both trials to increase the level of evidence. We will use individual patient data meta analytic methods, e.g. using the amount of Bias as weighting factor and compare this to other approaches described below. The approach is beyond the recommendation of the Cochrane Collaboration, where the evaluation of bias is on a qualitative level. Bias may cause potential unobserved heterogeneity which could be detected with the method at hand.
3. Further, within the IDEAL project some innovative statistical methodologies were developed to compare two or more groups [references] which are useful to reanalyze the data of the EPISTOP in different directions and with a possible increase in statistical evidence from the data. These groups can either correspond to two treatments in one clinical trial with the goal to detect differences, or they can correspond to different clinical studies with the goal to combine the data from different resources. In the project we will demonstrate the new methodology in both situations. We will compare the effect of medication in the two treatment groups (early versus late treatment) in a better way using statistical tests for comparing curves and surfaces. Starting with the primary endpoint (time to first seizure), we will compare the curves estimated from the Cox proportional hazard model (percentage of seizure-free patients and hazard function) corresponding to different groups. For this purpose, we will extend the maximal deviation approach developed in Bretz et al. (Bretz et al., 2018), Dette et al. (Dette et al., 2018), Collignon et al. (Collignon, Moellenhoff and Dette, 2019) and Möllenhoff et al. (Moellenhoff et al.,2018; Möllenhoff, Bretz and Dette, 2019), such that it can be applied to EPISTOP data. In a second step, secondary endpoints (such as time to clinical seizures, time with seizures, risk of hypsarrhythmia and infantile spasms, time to second drug and to seizure onset from first abnormal EEG) will be investigated. A particular focus in these considerations is the inclusion of covariates (such as tuber volume, RML volume, age etc.), which results in the comparison of surfaces. Similar clinical and EEG data were gathered in different centers of the EPISTOP consortium for patients not included in the trial. The methodology developed in IDEAL will be used to validate, if these data can be included in the EPISTOP study to conduct statistical inference based on larger data sets.
4. To reanalyze response data to treatment, for the randomized as well as for the observational trial separate longitudinal models will be applied, including Cox regression models to model e.g. recurrent clinical seizure events to understand the disease progression. With respect to score like the EEG score longitudinal data modelling is biased by ceiled or floored effects of data. This can be accounted for in nonlinear mixed effects models which are currently applied to SARA score in Friedreich Ataxia patients (Reetz et al., 2016) (Hilgers, work in progress). The improved model building will be used to illustrate projections about sample size calculation similar to the work successfully applied in Friedrich Ataxia (Reetz et al., 2019).

Finally, we will use model averaging methods, bootstrap and permutation test methodologies in advanced statistical models to overcome the uncertainty caused by bias in measurements. As this methodology is underdeveloped right now, it combines the research of the two IDEAL expert groups leaded by Professor Dette and Professor Hilgers.