We have an *opening for a PhD student* in my group on machine learning generalization "out-of-table". Help build methods that learn from large volumes of tabular data to generate models for new tasks! Apply here:

This is a great chance to study and influence the impact of medical AI applications, which are rolling out in our hospital. 🤖 🏥 Please help spread the word! If it could be you, check out the vacancy! Do reach out to me for any question.

📣Vacancy alert! 📣 We are hiring a *PostDoc in Causal Inference* for a project at the Amsterdam University Medical Center. Details here: https://t.co/FrqmmBPH2j

The climate issue alone is bigger than any other discussion point, even if we do not fully see its consequences yet. Consider voting for Harris, or abstaining at the very least.

The US has massive influence: if it delays taking action then the whole world will suffer. Trump's agenda on the climate theme will make the problem worse. If you intend to vote for him, please take a moment to consider the damage he might do.

Just this year, we had devastating floods in Europe, destroying large parts of Spain, Italy, Austria and many other countries. These are not fake reports, I know many people that were affected. Similar events are happening around the world and the planet is steadily warming up.

Dear American friends and colleagues, Please take a second to read this short thread. Your vote tomorrow matters. Regardless of your political beliefs, you must acknowledge that we are in the middle of a climate catastrophe. *Your vote can determine how bad it will be*

Many thanks to collaborators Stefano Finazzi and Frederik Hessulf! https://t.co/9IwaWwJiGF

Happy to share that our project TRAIL-ICU, has been awarded the AiNed XS Europa grant! The project aims to establish international collaborations in the realm of AI for the ICU, connecting researchers at @amsterdamumc, @MarioNegriIRCCS and @sahlgrenska University Hospital.

Curious to know more? Today an online interview was released where we answered in an accessible way to some questions about our project. Enjoy the reading! #XAI #NLProc #LLMs https://t.co/rm2UPCe3KS

A few months ago, @sandropezzelle and I were awarded a grant by the University of Amsterdam. We proposed a method to bridge #AI representations to human-readable explanations. Our method can be applied to #NLP, #ComputerVision, and #aiinhealthcare .

Great to see the TRIPOD+AI reporting guideline published! It does medical research a great service, but prognosis models intended to support treatment decisions need more, so we wrote a rapid response in the BMJ: https://t.co/LXRL08UWRc

"Dear friends, scientifically, this is not a #ClimateCrisis. We are now facing something deeper. Mass extinction. Air pollution. Undermining ecosystem functions. Really putting humanity’s future at risk. This is a #PlanetaryCrisis." — Professor @jrockstrom #ActOnClimate

New study: the Atlantic overturning circulation AMOC “is on tipping course”. The paper by Dutch colleagues adds more weight to recent warnings, such as the OECD Climate Tipping Points report of 2022 and the Global Tipping Points report published 2023. https://t.co/yJjUGRlqXm

This paper has been an eye-opener for us and we hope it will spark more discussion on the deployment and monitoring of prediction-based treatment policies. Finally, many thanks to wonderful co-authors @WvanAmsterdam @GelovenNan @jessekrijthe Rajesh Ranganath for their help!

3. Monitoring discrimination prospectively after deployment will not reveal whether the model is causing harm or not. In some cases, good calibration may also not say much (see our last theorem). We need better metrics to monitor decision support tools in production.

2. Implementing a personalized outcome prediction model is not always beneficial, even if a) the model is very accurate on historical data and b) the treatment in question has positive average treatment effect on the whole population.

Some common practices need revision. 1. Developing models on observational data ignoring the historical treatment policy is potentially dangerous for decision making, as the change in treatment policy pre-/post-deployment is what determines the effect on patient outcomes.

Our results define the conditions leading to a prediction-based policy being harmful and self-fulfilling, and also show that calibration may not be a good monitoring metric. The paper also presents a medically-realistic example of such a scenario.

This issue is not new in the Medical AI literature (e.g. Obermeyer 2019 paper), but our contribution is to lift the anecdotes to a theory allowing for the precise characterization of harmful self-fulfilling policies.