This week, an international audience set off to join the TrustworthyAI4Health workshop at EMBL. Over 70 healthcare and AI experts met to listen to compelling keynotes and flash talks, network during the poster exhibition, and develop joint ideas in the four discussion sessions. Here’s a little recap of the day.
Good morning – It’s keynote time
The organisers and hosts, Antti Honkela, Hakime Öztürk, and Oliver Stegle, opened the event with a warm welcome and a brief overview of organisational matters.
Four keynotes set the theme and topics for the day.
“Trustworthy AI in the Clinic: Bringing the Human Back into the Loop”
Jean Louis Raisaro analysed how research translates into everyday hospital work. He presented two research projects, a sepsis detection application and an LLM for the hospital, and analysed common challenges like trust or limited hospital infrastructure.
His key takeaway: Include clinicians early in your research. Let them evaluate your models. Collaboration is key.
Mennatallah El-Assady, ETH Zurich
Mennatallah El-Assady shared a deeper look at LLMs: Where do models learn to contextualise and disambiguate a token? What is the impact of white space, like an additional space, in a prompt on a model’s answer? How do you treat bias in data? Does your model need to know that more secretaries are female? What about the gender distribution of a disease?
“Robust and uncertainty-aware AI systems for high-stakes applications”
Uncertainty and confidence of algorithms were the main topics of Florian Buettner’s keynote. While focusing on algorithms rather than applications, he discussed the impact of image distortions on predictions in melanoma detection, the training and auditing of AI monitoring tools, and the efficiency of repeatedly asking a model the same question.
“Curious findings about medical image datasets”
How does medical imaging work? Why should we be sceptical? And which shortcuts the AI models could take do we need to be aware of? – These are only a few questions Veronika Cheplygina posed in her keynote. She raised awareness for critical thinking and why high success rates should trigger researchers’ scepticism. Spoiler: According to a certain model, it could still detect lung collapses when researchers hid the lung on the X-ray it used for detection.
Quick – Flash Talks
The participants also had the chance to listen to the quick talks by Anıl Kuş (Yozgat Bozok University), Arnisa Fazla (Amsterdam UMC Doctoral School), Farah Briki (CHUV) and Rajesh Divakaran (United Nations High Commissioner for Refugees ).
Thank you for sharing your research and projects!
Now together – Discussion Sessions and Panel
After choosing their preferred topic, the day continued for our participants with joining one of four discussion sessions:
- What would it take to shift the field towards routinely reporting and optimising across multiple metrics?
- How should we balance the need for well-calibrated, reliable uncertainty estimates with the demand for end-to-end interpretability in clinical AI systems?
- Which explainability perspectives do we need in order to get the right guarantees for safe medical or clinical use?
- How do we protect and model data privacy without degrading the reliability, contectual accuracy, and clinical usefulness of AI systems deployed in care?
In the sessions, the facilitators and keynote speakers framed the topic. Then, the participants had 90 minutes to form subgroups. They went into deep-dive discussions, created a poster with their insights, arguments, challenges, ideas, and solutions, and presented them to the rest of the group.
At the end of the day, participants, keynote speakers, organizers, and an additional panelist came together again for the final agenda item: the panel discussion.
Moderated by Oliver Stegle and enlivened by audience questions, the group engaged in a lively discussion.
And that was it – a successful workshop ended. Thank you to the organisers, speakers, panelists, flash talkers, poster presenters, and to our participants. You made this day.
A more technical summary of the day will be published soon. Until then, have a look at the event website for more info on the day and the posters: event website.