At UT, Alexia Briassouli builds smarter systems that help algorithms see, learn, and support better care.
Alexia teaches AI to identify meaningful changes, such as the early signs of a tumour or the moment a routine becomes a risk. But beneath that, she is decoding behaviour, patterns of health, and changes in them, which sometimes can make the difference between catching something early or too late. “We are helping real people, that is the heart of it”, she says.
Patterns and predictions
She began in signal processing, learning how to pull meaning from noisy data. But over time, the questions shifted: What if this change in movement patterns indicates that someone is at risk of falling at home? What if this image is a scan that shows something subtle but serious?
In one early EU project, she worked on technologies to support people with dementia living independently at home. This was more than a decade ago, before wearable tech went mainstream. “We had cameras, sensors, smart plugs, even wristbands that could detect physiological measurements like heart rate and sweat levels. We were already thinking about AI and health before the AI boom,” she recalls. Her work has always followed the same idea: notice the pattern, and then notice when it breaks.
Making healthcare smarter
Alexia’s work ties directly into UT’s mission to make healthcare smarter, more personalised, and more preventative. She works across disciplines (mathematics, computer science, engineering) but with one purpose: to create tools that doctors and patients can use. She has supervised dozens of student projects, from tumour detection and localisation models to finding where and why AI systems make mistakes. “In the end, it is always about noticing what has changed,” she says. “And asking why.” Even in public spaces, she found parallels to healthcare: tracking how people’s movements change around an artwork is not that different from finding where patterns change in medical images, or tracking a patient’s recovery.
Real-world AI has messy edges
Alexia now leads projects on robust AI in medical imaging, helping systems hold up in the unpredictable reality of hospitals, even when real life gets in the way. AI models are often trained on clean, ideal data: people who sit upright, stay still, and fit neatly into categories, with data from the same scanner. But real-world data often simply does not match the data the model was trained on: some patients may be in wheelchairs, some move unpredictably, and data can be from a different patient cohort or recorded by a different scanner.
“AI is as good as the training data,” she says. “If it has not seen enough variation, it starts making bad guesses.” Or, as she puts it: “AI loves to take shortcuts.” If it can rely on a lazy rule or a visual clue that worked before, it will. Even if that shortcut fails in a new situation. To fix this, her team builds smarter training sets, in some cases even creating synthetic scans that mimic edge cases (such as scenarios where patients move differently, or where data is incomplete), and develops models that rely on more robust, general features, for reliable, interpretable AI. The goal is that algorithms do not get confused when something unexpected happens.
Alexia wants AI systems that doctors can trust. Not just because they are accurate, but because they make sense. “It is not enough for a model to be right,” she says. “You need to understand why it is right, and know it will still work when things look a little different.” That is why her team focuses on building AI that performs well in real-world situations and can clearly explain its choices.
The importance of data: from the Olympics to hospitals
To explain why good data matters, Alexia tells a story from outside the hospital. The US Olympic swimming team, ahead of the 2024 games, worked with data scientists to optimise every movement. They collected large amounts of multi-sensor data and analysed it using data science, physics and biomechanics. As a result, they predicted, among others, that by improving head position during her underwater breaststroke pullout, the US breaststroke swimmer Kate Douglass could reduce drag and shave off up to 0.15 seconds from her pullout. Indeed, in the 2024 Paris Olympics, Douglass improved her 200-meter breaststroke time from 2:30 to 2:19.30, setting an American record and securing a gold medal. And demonstrating the power of good data and data science.
“That is what the right data and sophisticated data analytics can do,” she says. “For athletes, patients, and everyday people, it is the same: data helps you see where you are and what might happen next.” Alexia’s focus is on what AI can reveal, support, and make possible. “I want the systems we build to work for people. To help them live longer, stay safer, and feel more seen. That is the point.” And maybe that is the difference between teaching an AI to see and teaching it to care.
