Two case studies will be presented to demonstrate how patient preference studies can support decision-making in early drug development (Phases 1-2). The case studies used two different quantitative types of preference study: a Discrete Choice Experiment and a Swing Weighting (SW) approach.

The first case study used a DCE and obtained preferences from patients suffering from Chronic Obstructive Pulmonary Disease (COPD) [1]. It involved a total of 1050 patients from five countries and aimed to understand the relative importance to patients of different endpoints and to inform the choice of clinical, PRO, and digital endpoints for inclusion in pivotal clinical trials in COPD.

The second case study [2], which used the SW approach, is a patient preference study in a rare chronic kidney disease, IgA Nephropathy. Here the interview-based SW methodology was used to enable robust preference data to be derived from a small number of patients. The results of the study were used to evaluate the relative importance to patients of different benefits and risks, including the trade-offs that patients are willing to make.

References

1. Cook, N.S., Criner, G.J., Burgel, P-R, Mycock, M., Gardener, T., Mellor, P. Hallworth, P., Sully, K., Tatlock, S., Klein, B., Jones, B., Le Rouzic, O., Adams, K., Phillips, K., McKevitt, M. Toyama, K. and Gutzwiller, F. (2022). People living with moderate-to-severe COPD prefer improvement of daily symptoms over the improvement of exacerbations: a multicountry patient preference study. ERJ Open Research, 8(2):686-2021. DOI: 10.1183/23120541.00686-2021

2. Marsh, K., Ho, K-A., Lo, R., Zaour, N., George, A.T. and Cook, N.S. (2021). Assessing Patient Preferences in Rare Diseases: Direct Preference Elicitation in the Rare Chronic Kidney Disease, Immunoglobulin A Nephropathy. The Patient, 14(6), 837-847. doi: 10.1007/s40271-021-00521-3.

The prediction of drug concentration time courses after different dosing scenarios is greatly facilitated if the pharmacokinetics (PK) can be assumed linear. The assumption of linear PK thus needs careful evaluation for any new drug in development. Under linear PK, exposure is proportional to dose (linear PK across doses) and exposure at steady state can be predicted from a single dose based on the superposition principle (linear PK over time). While investigation of dose-proportionality is common practice, evaluation of time dependent PK has received less attention in the literature. In particular, the superposition principle can be used to assess whether the observed extent of accumulation after repeated administration is expected under the premise of linear PK. This work emphasizes the importance of the time related aspect of linear PK by introducing the predictability ratio (PR). Linear PK over time can be concluded if PR = 1. Accumulation is higher than expected if PR >1, and lower if PR <1. If PK data from multiple dose cohorts are available, the PR is assessed for each dose cohort and a supportive hypothesis test can be applied to test for potential differences between doses in PR.

In Biosimilar development, a human PK (PD) clinical pharmacology study is an essential step in the stepwise approach for demonstrating biosimilarity. Such a study is commonly designed to demonstrate 3-way PK (PD) similarity between a Biosimilar candidate, a US approved reference and an EU approved reference in multiple coprimary endpoints (e.g., AUCinf, AUClast, Cmax) with high confidence. For each of the coprimary endpoints the null hypothesis of no-similarity is tested based on Two One-Sided test (TOST) for all 3 pairwise comparisons with respect to the similarity margins. In the calculation of the required study sample size, unknown expected treatment differences and covariance parameters of the coprimary endpoints are usually substituted by point estimates (guestimates) from historical data whenever available, while ignoring their uncertainty. However, misspecification of these parameters is likely to result in an underpowered study and hence a lower actual trial probability of success.

For bioequivalence studies with one primary endpoint, Ring et al. (2019) proposed the use of the assurance approach to account for uncertainty on unknown treatment difference, especially using the expectation of the power function with respect to the prior distribution of the treatment difference.

In this presentation we show how to extend the assurance approach for multivariate coprimary endpoints for testing three-way similarity by estimating the posterior distribution of the study power as a function of the vector of the treatment differences and the covariance matrix using multivariate normal and Wishard prior distributions, respectively. In addition, we discuss how to inform the prior distributions of the unknow parameters for the following scenarios: 1) Historical PK (PD) data available for all three arms; 2) Historical available in one reference arm; and 3) Historical is not available. We also implemented the approach using an R-shiny application.

Real-world evidence (RWE), a key part of integrated evidence framework, plays an increasingly important role in optimizing trial design and supporting the approval of new drug development, especially for orphan and oncology drugs. However, the use of RWE in biosimilar development is still exploratory and lacks clear guidance from health authorities. In biosimilar study design, RWE has potential implications for filling historical knowledge gaps to improve study design efficiency, i.e. improving/validating design parameters such as margins or variability; contextualizing study results in the landscape of marketed biosimilar drugs; determining design parameters when historical data for specific endpoints or populations, etc. are not available. A study was initiated to verify the reliability of RWE use as a source of historical knowledge.

This study used AACR GENIE (American Association for Cancer Research, Tumor Information Exchange for Genomics Evidence) real-world data (RWD) to evaluate the treatment effect of immunotherapy compared with chemotherapy in non-small cell lung cancer (NSCLC) patients. Patients were screened based on several criteria, including age, diagnosis, line of treatment (i.e. first line), setting (i.e. monotherapy), etc., to represent a similar population as used in a pivotal Phase III clinical trial.

Patients in the immunotherapy group were matched with those in the chemotherapy group by age, sex, smoking history, and histology based on propensity scores calculated by logistic regression models. Best responses during treatment based on imaging records and medical records were derived, respectively. Results for differences in treatment rwRR (real world response rates), rwPFS (real world progression-free survival), and rwOS (real world overall survival) were compared with those reported in clinical trial publications.

The GENIE database included 1849 patients as of September 2020. Out of the patients who met inclusion and exclusion criteria, 34 and 189 patients received fist line immunotherapy and chemotherapy, respectively. The treatment duration (months) was comparable between immunotherapy versus chemotherapy (Median [Q1, Q3]: 4.17 [2.79, 8.71] vs. 3.06 [1.91, 4.04]). The rwRR based on overall imaging documentation was 35.5% (11/31) for immunotherapy and 19.4% (6/31) for chemotherapy. The median rwPFS was 5.6 month (95% CI: 4.0 – 27.9) versus 8.9 month (95% CI: 5.6 – NA) for immunotherapy versus chemotherapy. The median rwOS was 26.3 months (95% CI: 19.7 – NA) with immunotherapy versus 19.6 months (95% CI: 11.5 – 33.9) with chemotherapy.

This study explored a method to assess the feasibility of applying RWE in biosimilar development. Similar search by involving additional real-world database like Flatiron and/or ConcertAI has also been initiated to provide more insight into the applicability of RWE for designing biosimilar trials. This part of results will be available during presentation.

Quantification Of Dataset Similarity For Small Sample Sizes:

In clinical studies, there are several situations where the research question requires identifying similarities between two datasets. For instance, when the aim is to improve the prediction model for a specific data site with limited observations, one approach is to include weighted data from external sites based on their similarity while maintaining the original distribution of the target data site. To calculate the weights and thus to quantify the similarity, one can make use of the inverse probability of belonging to the target site using logistic regression. However, the similarity may be underestimated by this approach if the target site has significantly fewer observations than the external even if the target and the external dataset are samples of the same population. To address this problem, we propose oversampling techniques.

Specifically, we have developed an iterative process where we continue to oversample the target site observations until we observe that the distributions of the target data, before and after including weighted data, remain the same. For comparing the distributions, we use Kullback-Leibler divergence and parametric methods. By carefully monitoring the distribution of the target data, we can avoid introducing bias while still increasing the sample size based on similarity. We evaluate the effectiveness of our proposed method using a simulation study under two scenarios: one with similar observations in two data sets and one with dissimilar observations with varying degrees of similarity. We compare the results obtained with and without oversampling and assess the impact of oversampling techniques on the performance in terms of prediction performance. Results indicate that oversampling the small target data can improve the quantification of similarity for obtaining weights in a prediction model, resulting in higher weights when the distributions of datasets are similar but not overestimating the weights when the datasets are dissimilar.

We demonstrate our approach using the International Stroke Trial (IST) data, including patients with acute stroke in different countries. The aim is to include weighted data from other countries to a country with limited data, while the distribution of target data plus weighted external data remains similar.

(This presentation is a joint presentation with the GMDS conference, and will also be presented at that conference with a different focus.)