"What is so mysterious about modern statistics?"
Magnus Boman is an AI professor at the Department of Medicine, Solna, at KI, and part of the AI@KI project – a strategic project initiated by KI's rector with the goal of collecting and describing all activities at KI related to artificial intelligence (AI). In addition, since 2023, Magnus has been involved in AI issues at MedTechLabs and as an advisor within Clinicum for research methodology involving machine learning.
Hello Magnus! Why are you a part of Clinicum's counseling network?
We get a lot of questions about AI and machine learning and Clinicum needs to grow in that area. Me and Max Gordon at Danderyd are currently the only outspoken method advisors in that district. We look forward to several colleagues feeling ready to chip in. More personally, I learn an extreme amount by listening to everyone's problems. On the meta-level, I also better understand the confusion of colleagues: what is so mysterious about modern statistics anyway?
What does an AI professor do?
Does a lot of programming and math, although there never seems to be enough time for either. Education, further training and the beautiful word skills training are also included of course. But most of the time is spent answering questions like "Is this easy or hard with AI?" or "I have a lot of unique data that should be usable for prediction or classification, how do I know if it's the right data and enough for AI?".
I myself am directly involved in about a dozen research projects as well, where I have different levels of competence, ranging from "almost competent" to "absolutely no competence". So I try to cobble together what I can actually do, which is usually some kind of self-learning model, if the project manager wants it. I rarely do anything on my own initiative, but instead try to meet the needs that actually exist.
What kind of machine learning or AI questions might you be asked?
All questions related to learning models! (There are certainly stupid questions in the field, but no one knows which ones are real, it's day note on that.) Most of the questions I get do not concern generative AI, but discriminative AI, i.e. predictions or classifications. Something that can be done with standard statistical methods, such as regression, but which has a "twist" of some kind, non-linearity or very high complexity.
How can I best prepare for a consultation with you?
It is usually enough to fill in the Clinicum form. But anyone who has read or heard something fantastically exciting can always send me links. Same with own work from those who need method advice I should read - I love how much I learn that way!
Magnus Boman's best AI tips for researchers:
- There are many myths surrounding machine learning models, for example, “Random Forest performs best,”. At Clinicum, we often reveal that, for some reason, such myths never hold true for your data.
- To compare quantitative outcomes (e.g., AUC-ROC) of different prediction models is almost always like comparing apples and pears. We can help you understand why.
- There are no threshold values that determine the clinical relevance of a model. We can explain how to build an impeccable model with learning and self-improving capabilities. You then make the decision whether it’s worth the effort—or if a regression model will do in the end.
- Fill out a Clinicum form and ask me your specific question.