Teaching AI to doubt: improving reliability in medical AI
© 2026 EPFL
In medical artificial intelligence, the most critical risk is often not incorrect predictions, but overconfident ones. Aurel Davy Tchokponhoue is working on a key question: how can AI systems better recognize and communicate uncertainty in high-stakes contexts such as breast cancer diagnosis?
His research focuses on classifying breast cancer subtypes using gene expression data. While current models can achieve high accuracy, they always produce a prediction — even when the data is unfamiliar or unreliable, which can lead to overconfident errors in real clinical settings.
The issue isn’t that AI gets everything wrong. It’s that it can be very confident when it shouldn’t be.
Rather than focusing only on accuracy, the work studies uncertainty quantification, comparing different approaches (deterministic, Bayesian, ensemble) to assess how reliably models can signal when they are unsure. The aim is to support clinicians by indicating when a prediction should be trusted or reviewed.
To test robustness, the researcher also developed synthetic data methods (GMGS1 and GMGS2) that simulate realistic but unseen clinical scenarios, allowing models to be evaluated under changing conditions.
Overall, the research points toward AI systems that not only predict, but also know when to abstain — making medical decision support more cautious and reliable.
Read the full article on Times of UM6P
100 PhDs for Africa is a programme stemming from the “Excellence in Africa” initiative developed by UM6P and the EXAF Centre (EPFL).