Interview Foundation Models in Medical Research
An interview with Prof. Fabian Theis, Head of the Computational Health Center, and Prof. Bjoern Eskofier, Group Leader for Translational Digital Health. The researchers hosted the first Munich Health Foundation Model Symposium on April 10, 2024, at the Helmholtz Munich Campus.
An interview with Prof. Fabian Theis, Head of the Computational Health Center, and Prof. Bjoern Eskofier, Group Leader for Translational Digital Health. The researchers hosted the first Munich Health Foundation Model Symposium on April 10, 2024, at the Helmholtz Munich Campus.
Foundation models are large-scale machine learning models that are pre-trained on large amounts of diverse data. They serve as building blocks for various AI tasks, enabling efficient transfer learning and fine-tuning for specific applications. Foundation models have revolutionized the field of AI by providing powerful, generalized representations of data, language, and knowledge. One prominent example of a foundation model is ChatGPT. Further specialized models have the potential to advance our understanding of complex diseases.
Why are foundation models important for medical research? Can they autonomously generate new insights?
FT: There is a lot of excitement currently about foundation models, not the least because of the dominance they have shown in natural language processing. After all, if you want to write a state-of-the-art chatbot nowadays, you absolutely need to have a large-language model (i.e. a language foundation model) in the background.
Similarly, it is expected that also in biomedical research a representation of many data in a single model could help boost particular downstream problems tremendously, and first papers have started to show that this can be the case. For me, the most exciting aspect will be to see how to deal with multimodal and multiscale data so common in biomedicine.
Whether models actually generate insights is hard to say. After all, they 'just' help in improving downstream predictions. However, the first papers show that they may learn intrinsically non-trivial facts about the underlying biology, such as gene regulation in the case of transcriptomics data sets.
TUM | © Astrid Eckert"For me, the most exciting aspect will be to see how to deal with multimodal and multiscale data so common in biomedicine."
Prof. Fabian Theis
What potentials do foundation models offer in the future for prevention and precautionary measures for people?
BE: Humans are individuals, and most are interested in healthy lives. However, our current healthcare system is not well designed for these matters of fact: it is based on population averages rather than personalization and it is reactive rather than preventive. Foundation models in healthcare promise new opportunities because they can individually tailor prevention and health pathways on a broad basis of genetic, environmental, and lifestyle data.
Copyright: Kurt Fuchs"Foundation models in healthcare promise new opportunities because they can individually tailor prevention and health pathways on a broad basis of genetic, environmental, and lifestyle data."
Prof. Bjoern Eskofier
Can foundation models help us identify more effective treatment approaches for specific diseases?
FT: For sure they can help us improve diagnostics already now, simply by seeing more data, more patients, and more views of the same patient within a single model. Whether this translates into actual more effective treatments is not solved. After all, for this one would not only need to predict a disease state but also the causal effect of a perturbation. We and others have first results that this may be true for cell-based drug screens, but this is not yet working on the organism level, so early times.
What developments and trends shape the future of medical research with foundation models?
BE: Foundation models can only be successfully trained on a broad basis of a large amount of individuals’ health data. This does not only include what we currently conceive as medical data (like medical images, biomarkers, and genetics), but also environmental and lifestyle that shape the 'personal health dataspaces', also called 'digital health twin' of a person. I am thus personally quite interested in the developments of such personal health dataspaces, including the current initiative of the European Union towards the European Health Dataspace (EHDS).
How can we accelerate the deployment of foundation models in practice, and what prerequisites need to be established for this?
FT: Since first such models are only coming out now, deployment is mostly happening in research. This could be from just downloading the model and running it on your own data to using some online service to avoid sometimes infeasible computational costs locally. I would however expect soon that first such results will be finding their way into clinics, given the promising results as well as the excitement and already now significant investment of venture capitalists into startups with related missions.
Latest update: April 2024.
