Modern medicine increasingly relies on high-dimensional data - genomic, proteomic, imaging, and continuous physiological signals from clinical systems and wearables - that must be turned into reliable, individualized insight. The MedLab develops mathematical and machine-learning methods that make this possible: models that are not only accurate but interpretable, so that clinicians and patients can trust and act on their results. This spans the identification of multivariate disease signatures across -omics and imaging data, the construction of multi-scale models that embed these signatures in their biological context, and the extension of point-wise clinical analysis toward continuous monitoring of large mobile-health data streams - enabling earlier diagnosis, better therapy monitoring, and real-time tools for high-risk patients.
Safe and interpretable AI for biomedical applications
Artificial intelligence is transforming healthcare and the adjacent biomedical and medtech industries - raising both the potential and the expectations for what data-driven methods can deliver. This dynamic demands innovative approaches not only for developing AI-based methods for critical, sensitive domains, but also for continuously validating and monitoring them once deployed. Building on two prior funding phases, in which the MedLab developed methods for capturing and exploiting biomedical data and established the technological infrastructure needed to analyze very large datasets, the third phase pursues two goals in parallel: developing AI-based services for concrete biomedical use cases with our industry partners, and building a platform for the continuous validation, monitoring, and certification-readiness of these AI systems - increasingly relevant in light of the EU AI Act's requirements for high-risk AI systems.
Projects
In the third phase of the Research Campus MODAL, the MedLab is working on four sub-projects, three developing AI-based services with industry partners and one building the shared validation platform they integrate into.
AI-Based ECG Reporting
In this project, we develop an automated system that generates free-text diagnostic reports from ECG data, addressing the slowness and inconsistency of manual reporting. A ResNet-based encoder transforms ECG signals into embeddings, which advanced language models (LSTM/Transformer, including LLMs) turn into natural-language diagnostic text — evaluated against physician reference reports for clinical relevance.
AI-Based Anatomical Reconstruction
Here, we develop methods for reconstructing detailed 3D anatomical models from sparse or low-radiation 2D imaging data, using category-level statistical shape priors. A modified Gaussian Splatting Model generates virtual X-ray training data, used to train a Neural Radiance Field (NeRF) that reconstructs 3D anatomy from real and virtual 2D X-rays, with extensions for varying anatomical topologies.
AI-Based Image Segmentation
We build an end-to-end AI system to replace slow template-matching methods for identifying macromolecular structures in cryo-electron tomography data, where imaging speed has outpaced analysis capacity. The approach combines synthetic training data, real experimental tomography data, and domain-gap reduction techniques (image filtering, GAN-based data adaptation) to bridge synthetic and real data.
Platform for AI Validation, Evaluation and Monitoring
The unifying platform of the MedLab: it enables real-time performance and safety monitoring of deployed AI systems, distribution-shift detection, systematic validation including Explainable AI (XAI), and documentation support for regulatory certification (EU AI Act). The platform can run multiple AI approaches on the same task in parallel - for example, several ECG classification models evaluated simultaneously against expert-verified ground truth - to continuously compare and identify the most effective solution.
