Abraham Vaquero Castro 1, Enrico Grisan 1, Monica Simeoni 2
1 London South Bank University (London, United Kingdom), 2 GSK (London, United Kingdom)
1. INTRODUCTION
Pharmacokinetics-Pharmacodynamics (PK/PD) data analysis is a cornerstone of both drug development and efficacy and safety studies. However, individual-level PK/PD data are difficult to obtain, expensive, and scattered throughout different clinical trials, for which usually only aggregated statistics are publicly reported.
Meta-Analysis (MA) approaches from simple MA to the more advanced multi-variate meta regression, and Model-Based MA (MBMA) are among the available tools to interpret average-level data. Ideally, the availability of individual patient data (IPD) would allow methods based on parametric pharmacological models, to provide a better characterization of the relationships between covariates and PK/PD parameters. We propose to leverage a generative-AI approach to reconstruct the IPD of cohorts with only population-level statistics, by exploiting the availability of a small set of IPD.
Differently from previous research aiming to integrate Average level Data (AD) and IPD in a combined meta-analysis (S.Weber [1], L.Yang [2]), this project utilizes a new approach by exploiting a generative-AI algorithm to regenerate IPD from AD, with the aim of performing an IPD MBMA.
The case study presented was used to test and validate the methodology. Using IPD from 2 clinical trials for two different drugs indicated for COPD as well as an AD database that contains evidence for over 150 clinical trials and 30 compounds, from published sources, we reconstruct the missing information for the averaged data, resulting in a complete IPD dataset for all the trials. In order to test the quality of the reconstructed dataset, we then perform a MBMA using a variation of the model presented by (C.Llanos [3]), adjusted for IPD data and similar to the one used in [2].
2. OBJECTIVES
• To design an AI/ML pipeline to reconstruct published and averaged external evidence into Individual-Level datasets.
• To investigate goodness of fit for the AI/ML data by using an already validated MBMA model.
• To compare the results from our IPD MBMA to other analyses performed on either mixed AD and IPD, or AD only.
3. METHODS
1. AD collection: The list of papers from which we extracted the AD for the experiment was the same as in [3]. By doing this, we were able to obtain a baseline for what our final estimates should look like. For the experiment, we aimed to generate as many individual cohorts as participants were in the original trials, in this way sample sizes for the averaged distributions would stay relevant to our reconstructed dataset.
2. Reconstruction algorithm (Conditional Variational Neural ODE): Throughout the course of the project, different algorithms were tested, including Bayesian Neural Networks [4], Wasserstein GANs [5], Variational Auto-Encoders [6] and Variational Auto-Encoder Neural ODEs [7], having the latter as
the chosen model due to its performance and characteristics. The core structure of the model is that of a Conditional Variational Auto-Encoder, having the decoder replaced by a Neural ODE. The first part of the algorithm compresses the input evidence into the latent space, where the relationships between variables and data distributions are held. This information is then conditioned by an additional set of information, in our case this is the change from baseline values at different timepoints (collected on Methods 1.) as well as a simple EMAX model, which can be switched off in case we don’t have prior information for the studied compounds or we prefer a fully agnostic approach. Having our latent space ready, it is accessed by the Neural ODE, which is used to predict the FEV1 progression for each individual patient at any timepoints we declare.
3. MBMA: In order to perform the meta-analysis, we used a NONMEM model that combined the information of [2] and [3]. The parameter estimations were conducted by using First Order Conditional Estimation (FOCE) method and using the results from the previous papers as initial estimates. While on validation phase, we used a reduced version of the reconstructed dataset that contained 12077 patients and 132847 timepoints, which relates to cutting down the final dataset to a maximum of 25 patients per arm and trial. For computational purposes, the results in the following sections were obtained by validating with this reduced dataset.
4. RESULTS
Several visual and statistical tests were performed to ensure the goodness of the synthetic generated data. In the Machine Learning section, we studied the overall deviation of the predicted FEV1 progression from the reported change from baseline values, where available, indicating an average variation 6.3% over the 156 studies where the twin population was obtained. For the model fitting results, we compared the theta estimates obtained from executing the model from [2,3] with our reconstructed data. Looking into the drug effect parameters, we observe a 0.904 Spearman correlation value with both projects. In terms of the structural parameters for Baseline, Disease progression and Placebo models, we obtained a 0.927 correlation with [3] and 0.576 with [2]. Furthermore, we inspected not only the mean values for these parameters but also the RSE values. The estimates from our
methodology present a reduction in the uncertainty for all theta values and they are located within a 50% interval from the ones in the other methodologies when normalised to the estimates in [3]. We find an exception in two parameters, specifically the ones accounting for the effect of disease severity and age in baseline.
5. CONCLUSIONS
The proposed methodology shows that generating a twin population and obtaining the measurement predictions from the AI/ML model produce individual level with an accuracy that matches the average level data when used for MBMA. Whether MBMA produces better estimates when performed using IPD or AD is still a topic of further investigation in simulation/re-estimation scenarios. In our case study, we observed estimates obtained with lower RSE than in other projects where a combination of AD and IPD or AD only were used. The mean value of the estimates exhibits a sufficient level of correlation with the other methodologies to ensure the goodness of the generated data. The methodology has the potential to improve the understanding of covariate effects and patient population in drug development.
References:
[1] S. Weber, A. Gelman, D. Lee, M. Betancourt, A. Vehtari, and A. Racine-Poon, “Bayesian aggregation of
average data: An application in drug development,” The Annals of Applied Statistics, vol. 12, no. 3, Sep. 2018,
doi: 10.1214/17-AOAS1122.
[2] L.Yang, C.Llanos-Paez, S.Yang, C.Ambery, A.Berges, M. Kjellsson, M. O. Karlsson. “A combined model
meta-analysis of aggregated and individual FEV1 data from randomized COPD trials”. PAGE 32 (2024) Abstr
10848.
[3] C. Llanos-Paez, C. Ambery, S. Yang, M. Beerahee, E. L. Plan, and M. O. Karlsson, “Joint longitudinal
model-based meta-analysis of FEV1 and exacerbation rate in randomized COPD trials,” Journal of
Pharmacokinetics and Pharmacodynamics, vol. 50, no. 4, pp. 297–314, Aug. 2023, doi: 10.1007/s10928-023-
09853-z.
[4] M. Arjovsky, S. Chintala, and L. Bottou, “Wasserstein GAN”, Jan. 2017, doi: 10.48550/arXiv.1701.07875
[5] E. Goan, C. Fookes, “Bayesian Neural Networks: An introduction and Survey”, Jun. 2020, doi:
10.48550/arXiv.2006.12024.
[6] D.P. Kingma, M. Welling, “Auto-Encoding Variational Bayes”, Dec. 2013, doi:10.48550/arXiv.1312.6114.
[7] M.L.Garsdal, V.Søgaard, S.M. Sørensen, “Generative time series using Neural ODE in Variational
Autoencoders”, Jan. 20220, doi: 10.48550/arXiv.2201.04630.
7. DISCLOSURE STATEMENTS
• This work has been funded by GSK and London South Bank University.
• Monica Simeoni is an employee of GSK and holds financial equity in GSK
Reference: PAGE 34 (2026) Abstr 11934 [www.page-meeting.org/?abstract=11934]
Poster: Methodology - New Modelling Approaches