AIPD-DC13-FH

• Host Institution: Fraunhofer SCAI
• Country Host Institution: Germany
• PhD Enrolment: University of Bonn
• Start Date: October 2025
  
• Duration: 36 months
• Official PhD Supervisor: Holger Fröhlich

Research Objectives

In the past FH has developed generative AI approaches (MultiNODEs), which can be trained on longitudinal cohort studies and allow for a realistic simulation of synthetic patient trajectories. Briefly, MultiNODEs encode observed multimodal longitudinal data into the initial conditions of a latent ordinary differential equation system (ODE), which is parameterized on the right-hand-side via a multi-layer perceptron. MultiNODEs allow for simulation, interpolation and extrapolation of patient trajectories on a continuous time scale, and such simulations may be further constrained by observed characteristics/covariates of a specific patient, i.e. provide a “personalized” simulation of possible outcomes.

Using data from NCER-PD/LuxPARK, PPMI and ICEBERG, this project has the objectives to:

1. Implement a generative AI approach for “personalized” simulation of disease outcome trajectories (e.g. UPDRS sub-item scores). Notably such a personalization may also include information about patient strata or medications. Accordingly, counterfactual simulations (e.g. change of medication at study baseline) become possible by altering defined covariates. The personalized simulation may be understood as a digital PD twin.
2. Develop methods to better interpret the latent space dynamics learned by MultiNODEs, e.g. via sensitivity analysis. This will result in better explainability of the model.
3. Develop an approach to merge synthetic patient trajectories across studies, e.g. to generate a “global” synthetic control arm. This is principally possible, because MultiNODEs generate data on a continuous time scale. However, initial conditions may systematically differ between studies. The main task is thus to learn a mapping between the initial conditions from different studies.
4. Develop methods that allow a selection of synthetic controls according to user defined selection criteria. The approach will be tested with the help of published inclusion/exclusion criteria for PD trials on clinicaltrials.gov.

This project will collaborate with the project of DC7 due to the common focus on generative AI and synthetic data.

Expected Results

• Innovative AI/ML approach for generating (personalized) synthetic patient trajectories and for identification of fast progressing patients.
• Innovative AI/ML approach for merging synthetic data across studies and for generating a synthetic control arm.
• Careful assessment of potential impact on clinical trial design.
   

Planned Secondment(s)

• Host 1: Novo Nordisk
  
  • Duration: 18 months
    
  • Purpose: Learning about clinical trial design and simulation
• Host 2: RCSI University of Medicine & Health Sciences
  
  • Duration: 6 months
    
  • Purpose: Learning about PD neurology and biomarkers

This project is part of the "Digital Health" work package.