(M-083) Disease progression simulation via generative AI: assessing the credibility of a conditional longitudinal autoencoder via a case study in Alzheimer’s Disease

Nick Henscheid – Quantitative Medicine – Critical Path Institute; Jagdeep Podichetty – Quantitative Medicine – Critical Path Institute

Objectives: Clinical trial simulation tools rely on mathematical models for the natural progression of disease, typically fit to historical data. Traditionally, nonlinear mixed effects models have been employed for this purpose, but generative AI holds promise for modernizing such simulation tasks. While AI-based methods can discover complex nonlinear relationships between multimodal covariates and longitudinal outcomes, the credibility of such methods must be assessed with respect to the specific context of use. Our goal is to build a framework to assess credibility for such models, using Alzheimer’s Disease as a case study.

Methods: Historical trial data were curated from the Critical Path for Alzheimer’s Disease (CPAD) database (14 studies, N = 4428). A generative AI method was built based on the Conditional Variational Autoencoder (CVAE) framework [1], with baseline patient covariates used to inform the conditional prior model, and Long Short-Term Memory (LSTM) models employed for the encoder and decoder layers. An existing regulatory-grade Nonlinear Mixed Effects (NLME) model [2] was implemented in Stan for benchmarking. For both models, cross-validation splits were used to evaluate simulation performance, with Visual Predictive Checks (VPCs) and statistical metrics between real and simulated data used to assess performance. A comparison between the modeling approaches was then developed to assist in evaluating credibility with regards to model interpretability.

Results: Both the NLME and CVAE-based methods produce realistic simulation output using baseline covariates to predict longitudinal outcomes. As assessed by VPCs, the CVAE method moderately improves upon the NLME model’s simulation performance, particularly in later times where the NLME model tends to overestimate severity; technical performance metrics comparing real and simulated distributions at fixed time points confirm this. These technical evaluations in combination with comparisons between the interpretability of the CVAE and NLME frameworks show that it is possible to establish credibility for a generative AI-based disease progression simulation model.

Conclusions: Generative AI shows promise as a tool for simulating disease progression, but incorporating this new paradigm into existing simulation workflows requires careful consideration. Our goal is to demonstrate that the credibility of generative AI methods can be assessed in a way that highlights their strengths and provides decision-makers with the information necessary to determine their best usage. Future applications of this framework to additional disease areas and simulation targets are planned.

Citations: [1] Sohn, Kihyuk, Honglak Lee, and Xinchen Yan. "Learning structured output representation using deep conditional generative models." Advances in neural information processing systems 28 (2015).
[2] Conrado, Daniela J., et al. "An updated Alzheimer’s disease progression model: incorporating non-linearity, beta regression, and a third-level random effect in NONMEM." Journal of pharmacokinetics and pharmacodynamics 41 (2014): 581-598.

Keywords: Clinical trial simulation, disease progression modeling, generative AI