Reliable Health Intelligence Lab

Animation: trajectories of a simulated rhythm settle onto a stable cycle; groups pushed off course turn red and find their way back.

The window we can build

When the gods created the world, they asked Momus to judge their work. Athena had made a house, Poseidon a bull, and Prometheus a man. Momus criticized all three.

He said the house should have had wheels, so people could move away from bad neighbours.[1] He said the bull’s eyes should have been placed below its horns, so it could see what it was attacking. And he said the man should have had a window in his chest, so others could see his thoughts and know whether he really meant what he said.[2]

Zeus did not appreciate the criticism and threw Momus off Olympus.[1] His complaints were never really answered. They were simply dismissed.

Aristotle later responded to one of them: the criticism of the bull. Moving the horns, he argued, would affect the whole animal. A bull still has to eat, turn its head, keep its balance, and support its skull. Momus’s suggestion sounds obvious only because it ignores all the other limits the designer had to work with.[3] It is easy to criticize a design when you do not have to deal with the constraints behind it.

Both sides matter. A system whose claims cannot be checked is difficult to trust with real responsibility. But criticism is also limited if it ignores the practical constraints involved in building the system. Our lab works between these two ideas, but from the perspective of people trying to build things that work.

Modern AI models are unlikely to give us the kind of window Momus imagined. We may never be able to look inside a model and fully understand exactly how it reaches every decision. But medicine does not necessarily need that kind of window.

A clinician cannot fully see the inner workings of a patient either. What an experienced clinician can develop is a sense of when a patient is beginning to deteriorate.[4] That knowledge comes from evidence and experience. It does not explain every mechanism inside the patient, but it can still help someone recognize danger and act in time. That is the kind of window we want to build for AI.

Instead of trying to see exactly how a model thinks, we want to understand where it stops working reliably.[5] We want to know which patients or groups it performs poorly on.[6] We want to know which missing information makes its performance worse. We want to know how much change or pressure it can handle before it begins to fail, and whether those failures can be detected from the outside.[7][8]

This is not an argument for being more cautious with AI. In fact, the opposite can be true. Caution becomes necessary when we do not know where the limits are. When those limits are unclear, even useful systems can remain stuck in pilots for years because no one is confident enough to deploy them. That may look like safety, but it can also delay the use of systems that could help.

Medicine already has a way of dealing with this problem. Drugs are not prescribed because they are considered safe in every possible situation.[9] They are prescribed because we understand when they should and should not be used, and because monitoring systems exist to identify problems that clinical trials may have missed.[10] Knowing where something can fail is part of what makes it possible to use it safely.

We want the same approach for health AI: clear and limited claims about what a system can do, clear boundaries around where those claims apply,[8] and monitoring that is strong enough to detect problems after deployment.

With that kind of evidence, useful systems could be deployed earlier and with greater confidence. The main problem is therefore not simply regulation. The problem is that we often do not have enough evidence about where AI systems work, where they fail, and how those failures can be detected. Rather than only debating those questions, we want to produce the evidence needed to answer them.

That is the science the Reliable Health Intelligence Lab is here to do.

References (10)
  1. Aesop. Zeus, Prometheus, Athena, and Momus. Perry Index 100. English rendering in Aesopica. source
  2. Lucian. Hermotimus, section 20. Second century CE. English passage reproduced in The Mind Is a Metaphor, University of Virginia. source
  3. Aristotle. Parts of Animals, Book III, chapter 2. Translated by William Ogle. Internet Classics Archive, MIT. source
  4. Douw G, Schoonhoven L, Holwerda T, et al. Nurses’ worry or concern and early recognition of deteriorating patients on general wards in acute care hospitals: a systematic review. Critical Care. 2015;19:230. doi
  5. Ghassemi M, Oakden-Rayner L, Beam AL. The false hope of current approaches to explainable artificial intelligence in health care. The Lancet Digital Health. 2021;3(11):e745–e750. doi
  6. Subbaswamy A, Sahiner B, Petrick N, et al. A data-driven framework for identifying patient subgroups on which an AI/machine learning model may underperform. npj Digital Medicine. 2024;7:334. doi
  7. Finlayson SG, Subbaswamy A, Singh K, et al. The clinician and dataset shift in artificial intelligence. New England Journal of Medicine. 2021;385:283–286. doi
  8. International Medical Device Regulators Forum. Good machine learning practice for medical device development: guiding principles. IMDRF/AIML WG/N88 FINAL:2025. 29 January 2025. See principles 3, 6–10. source
  9. US Food and Drug Administration. Benefit-Risk Assessment for New Drug and Biological Products. Guidance for Industry. October 2023. source
  10. US Food and Drug Administration. Step 5: FDA Post-Market Drug Safety Monitoring. The Drug Development Process. Accessed 16 September 2026. source
the window Momus asked for, reimagined for an age of artificial intelligence

What we build and test

We develop machine-learning models for health data from everyday devices, and we evaluate how far they can be trusted once they leave the lab.

01 / signals

Wearable and voice biomarkers

Accelerometry, heart rate, PPG and body temperature from consumer wearables; voice and cough audio from phones. We turn them into measures of rhythm, ageing and disease.

02 / evidence

Robustness and failure evaluation

Finding the patient subgroups a model fails, testing missing data and shifts in devices or populations, counterfactual explanations, and signals that flag unreliable predictions.

Where it is used: home monitoring and chronic disease, including heart failure, respiratory disease, sleep and circadian health, and healthy ageing. We work with clinical partners and run our own studies on CLAID, our open-source data-collection platform.

Team

Portrait of Dr. Filipe Barata
Dr. Filipe BarataPrincipal investigatorDigital biomarkers from wearables and voice; reliability and failure detection in health AI.ETH profileScholarORCIDLinkedIn
Portrait of Arda Bariş Başaran
Arda Bariş BaşaranPhD studentMultimodal and vision-language models; a smart health assistant.ProfileEmail
Portrait of Mesut Ceylan
Mesut CeylanPhD studentMachine learning for digital biomarkers, cerebrospinal fluid analysis, and counterfactual explanations.ProfileLinkedInGitHub
Portrait of Prof. Elgar Fleisch
Prof. Elgar FleischProfessor, host chairProfessor of Information and Technology Management at ETH Zurich and the University of St. Gallen.ProfileLinkedIn

Join

We look for people who like breaking models as much as building them.

Students

We are always looking for motivated bachelor and master students to work on our topics. If you are interested in a thesis or project with us, send a short note with your CV and transcripts.

email fbarata@ethz.ch

Research and clinical partners

Have a model you want to deploy, or data you want to learn from? We design evaluation studies that show where a model works, where it fails, and how to monitor it once it is in use.

Send
  • The model or question and its intended use
  • The data available: signals, setting, population, size
  • Your timeline and what a good result would let you decide
email fbarata@ethz.ch