Biomathematics Seminar Series - Achille Fourtoy (Nov. 10, 2026)

Speaker: Achille Fourtoy  (Fraunhofer Institute for Scientific Computing and Algorithms, Novo Nordisk)

Title: Generative AI for Simulating a Digital Parkinson’s Disease (PD) Twin

Abstract: 

Background: Parkinson’s disease (PD) is the second most prevalent neurological disease and the fastest growing. As of now, no disease-modifying treatment exists, and its causes and contributing factors remain largely unknown. The AIPD project aims to develop advanced AI methods to improve understanding and treatment of PD. Randomized Controlled Trials (RCTs) remain the gold standard for clinical evidence, but are costly, slow, and difficult to recruit for, especially when longitudinal and multimodal data are required. This is particularly challenging in PD, where long untreated periods are ethically unacceptable, and underlying disease dynamics are yet to be understood. Synthetic control arms and medical digital twins (DTs) offer a promising strategy to emulate trials, de-risk trial design, and support personalized treatment, biomarker discovery, and patient stratification. 

Methods: We developed MultiNSDE, a latent neural stochastic differential equation (NSDE) framework capable of learning and generating highly heterogeneous data distributions in continuous time from irregular and incomplete clinical datasets. This includes static and longitudinal continuous, categorical, and survival outcomes. In particular, the model separates treatment dynamics from placebo dynamics, which enables treatment-informed trajectory simulation and extrapolation, as well as counterfactual generation in a causal framework. 

Results: MultiNSDE was evaluated on three neurodegenerative disease trial datasets: A4 in Alzheimer’s disease, PRO-ACT in amyotrophic lateral sclerosis, and DATATOP in Parkinson’s disease. The model reproduced static and longitudinal data distributions, extrapolated clinically plausible trajectories, and generated coherent counterfactual outcomes under treatment-to-placebo and placebo-to-treatment shifts. Synthetic data generated by MultiNSDE preserved the average treatment effect estimates observed in the real datasets. 

Conclusion: MultiNSDE provides a proof-of-concept framework for generating realistic mixed-type longitudinal clinical trial data while preserving clinically meaningful treatment-associated differences in outcomes. It may support disease modeling, counterfactual simulation, and synthetic control generation in neurodegenerative disease research. However, the model still lacks features required to reach patient-specific DTs, such as continuous data integration and patient-specific decision-making. Additionally, it needs to be evaluated on real-world and observational study data. 

"Funded by the European Union under the Marie Skłodowska-Curie grant agreement No 101168379. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor REA can be held responsible for them.”