Summary
I am a Group Leader at the MRC Biostatistics Unit, where I am establishing a research programme at the interface of statistics and machine learning. My current work focuses on biomedical applications, with particular interest in generative models for time-dependent data, latent dynamics through manifold learning, and the development of robust AI systems grounded in scientific, geometric and mathematical principles. Prior to joining the BSU, I was a Senior Research Scientist in the Biomedical Image Computing Group at ETH Zürich, I have also held research positions at the Institute of Astrophysics and obtained a doctorate in Computational and Theoretical Physics from ETH Zürich. I am currently a member of King's College, Cambridge.
An up to date list of my publications is available on my Google scholar profile. https://scholar.google.ch/citations?user=JUKtIdkAAAAJ&hl=en
Personal website https://k-flouris.github.io/
My research connects theoretical innovation in machine learning with pressing challenges in biomedicine and public health, with current focus on:
Generative AI: We develop machine learning models (e.g., flow-matching) that integrate scientific knowledge, like inspired by physical or biological laws, directly into their architecture or via data curation. This creates "mechanistic-inspired" AI that can generate realistic, interpretable data. For example, we use these models to discover hidden structures in complex data, such as integrating multi-modal imaging and omics data to understand spatial tissue heterogeneity.
Digital Twins: We aim to build patient-specific "Digital Twins" that simulate disease progression and response to treatment. By combining powerful generative models with established mechanistic models (e.g., tumor growth equations), we can predict patient outcomes in oncology and guide personalized medicine. This allows us to explore "what-if" scenarios for individual patients under different therapies.
Safety and Optimization of Clinical Trials: We use deep learning to simulate complex patient trajectories and potential safety issues in clinical development. This allows us to create "in-silico trials" to test and optimize the policies of adaptive clinical trials, helping to reduce patient risk and improve the efficiency of developing new drugs.
Dynamical Systems: We develop frameworks for making deep learning reliable for systems where the underlying mechanics matter, such as in epidemiology. This involves enforcing structural and physical constraints on AI models, ensuring they respect the rules of the system they are modeling. The goal is to produce robust and reliable estimates of key parameters, such as disease transmission rates, even when data is noisy or limited.
Medical Imaging: Our work involves extracting high-dimensional insights from complex, multi-modal data. We specialize in developing methods for learning compact latent representations and analyzing medical time-dependent data, which includes dynamic acquisitions like 4D MRI. This is crucial for modeling dynamic processes such as brain aneurysms and biomechanical kinematics.
Generative AI, Medical Imaging, Genomics, Epidemiology, Dynamical Systems, Structure-preserving deep learning
