Research
We develop machine learning methods for health data and put as much effort into validation as into modeling, so that results carry over to patients a model has never seen.
HIPPOCAMPUS · WHERE OUR ALZHEIMER'S MODELS LOOK
The big picture
How the lab's work fits together. Health data of every kind feeds knowledge graphs and models; agents put them to work; the result reaches clinics and patients, under the same rules of trustworthy AI throughout.
01Multimodal health data
02Knowledge & models
03Agentic AI
Multi-agent framework
04Applications & impact
✓Trustworthy AI & responsible innovation, at every stage
Machine learning on physiological time series: sleep staging from polysomnography, including newborns in intensive care, EEG-based detection of neurological and psychiatric conditions, and ECG classification.
Deep learning models for brain MRI and CT, with a focus on Alzheimer's disease, stroke and hemorrhage. We validate at the level of the patient rather than the image slice, so that slices from one person never appear in both training and test data.
Recognizing emotional and stress states from wearable sensors such as ECG and skin conductance, and auditing how the measurement hardware shapes the results.
Detection and staging of Alzheimer's disease from imaging and physiological data, aimed at earlier diagnosis and better care.
Large language models and multimodal frameworks that combine structured data, signals, images and text for clinical and public-health questions.
Machine learning on population-scale survey and registry data, for example identifying risk factors of substance use.
Illustrations generated with ChatGPT for the AI4Health Lab. They depict research themes, not study data.
How we work
In this illustrative sketch: forty people with six recordings each. A real study has thousands of 30-second sleep epochs or MRI slices per person, and its cohort is usually larger. No sample ever loses track of whose body it came from.
Split the samples at random and most patients end up on both sides. A model can then score well by recognizing a person's signal or anatomy rather than the disease, and the number drops on new patients or new sites. It is a common way for reported accuracy to be inflated.
Whole patients go to the test group, with the same share of diagnosed patients and controls as the full cohort. Every sample of a test patient is sealed away before any modeling starts.
Architecture, features, thresholds and preprocessing (such as normalization) are chosen and fitted on training patients only, by cross-validation: in each round one fold of patients checks a model trained on the others. The folds are split by person, too.
The final model is scored once, on the sealed patients. Eight patients give a wide margin of error, so with small cohorts we repeat the whole procedure with a different sealed group each round (nested cross-validation, where tuning happens only inside each round's training patients) and report the spread across rounds.
