Modern healthcare produces much data such as medical images, electronic records, and physiological signals, but these data are often incomplete or not aligned, limiting how artificial intelligence (AI) can use them. This project develops a new approach called test-time steered generative modeling, which lets AI learn from such imperfect data and adapt during use. The key idea is a paradigm shift: instead of training medical AI as a narrow input-to-output (X?Y) mapping, we model the joint distribution of multimodal data (X,Y) that can be steered for flexible inference. Hopefully this can be one step toward clinical world models. The research combines self-learning AI, generative modeling, and clinical data from lung cancer, heart imaging, and hospital records, carried out at ELLIS Institute Finland and Aalto University. The results will lead to more reliable, data-efficient AI tools that help doctors analyse complex medical data and support trustworthy clinical world models.

2026

2030