SmallData Seminar on usecases of synthetic data in biomedical research

Collecting well-annotated biomedical data is costly, slow, and often impossible at scale, for example in rare disease contexts where patient numbers are small and privacy rules are strict.

Synthetic data offers a way to fill gaps by enlarging cohorts or balancing under-represented subgroups, and testing counterfactual scenarios without exposing real patients.

This seminar introduces three scenarios of synthetic data for gene–disease discovery, wearable-sensor activity recognition, and longitudinal clinical modelling. Across these talks and subsequent discussions we will introduce pipelines, validation steps, and challenges in synthetic data utilisation.

Moderator: Clemens Schächter (A03)

Agenda:

• 13:00 – 13:15 Introduction into synthetic data and synthetic data generating techniques
• 13:15 – 13:45 Talk Adrian Fritz and discussion
• 13:45 – 14:15 Talk Lars Ole Häusler and discussion
• 14:15 – 14:30 Break
• 14:30 – 15:00 Talk Clemens Schächter and discussion
• 15:00 – 15:30 Closing remarks and open end discussion

Abstracts:

Adrian Fritz & Miriam Schmidts
The prediction of gene-disease associations relies on curated databases with experimentally verified associations, for example HumanNet and OMIM. While these databases steadily increase in size, deep learning methods could profit from having more data available. Other databases, like DisGeNet, may contain more genes or gene-disease associations but these introduce inferred or spurious associations. Models trained on these bigger datasets often show a worse performance in predicting the experimentally verified associations from OMIM. Synthetic data might alleviate this problem, but it has to be created carefully: Further associations have to be realistic as to not introduce more noise to the model.

Lars Ole Häusler & Oliver Amft
The presentation investigates the generation and application of synthetic Inertial Measurement Unit (IMU) data in Human Activity Recognition (HAR). We will explore multiple approaches, including modality transfer as well as text-based generative models, to create 4D human surface models performing various activities. The surface models enable us to synthesise IMU data at arbitrary positions, which simulates diverse sensor placements.
We will further discuss the feasibility of training HAR models solely using synthetic data with tests on public datasets, compare synthetic and real data, and examine the necessary steps to generate effective synthetic IMU data for deep learning models.

Clemens Schächter & Harald Binder
We use literature-derived, synthetic longitudinal trajectories for spinal muscular atrophy to span the latent space of a generative deep-learning model and make the latent space interpretable. Specifically, we use pairwise comparisons between real and model-generated data and a small set of trajectories that represent prototypical disease courses. Here, a large-language model acts as a synthetic clinical expert, classifying each sample as more similar to one trajectory or another. Afterwards, these labels are used during training: patients judged similar are pulled together in latent space, while dissimilar ones are pushed apart. The resulting regularisation disentangles densely populated regions with the aim of sharpening the resolution of treatment-switch effects without requiring manual expert annotation.