Public defence: Rabindra Khadka
Rabindra Khadka will defend his thesis “Structured Representations for EEG-Based Dementia Prediction: From Spatiotemporal Transformers to Multimodal Biomarker Regression” for PhD in Engineering Science.
Rabindra Khadka will defend his thesis “Structured Representations for EEG-Based Dementia Prediction: From Spatiotemporal Transformers to Multimodal Biomarker Regression” for PhD in Engineering Science.
The dissertation is available in NVA Digital Archive (nva.sikt.no).
Leader of the public defense: Siw Fosstenløkken, Professor, Department of Computer Science, OsloMet.
Main supervisor: Anis Yazidi, Professor, Department of Computer Science, OsloMet.
Co-supervisors:
Dementia develops slowly and silently over many years. By the time symptoms appear, much of the brain’s structure and function has already changed. Detecting these changes early, when treatment and prevention are still possible, remains one of today’s greatest medical challenges.
This work explores how artificial intelligence (AI) can analyse brain signals to identify the earliest signs of dementia. We focus on electroencephalography (EEG), a non-invasive and affordable method that measures the brain’s electrical activity through small sensors placed on the scalp. EEG is widely used in hospitals, but interpreting its complex wave patterns is difficult and often depends on specialist expertise.
The goal of this research is to develop machine learning frameworks that automatically extract and interpret meaningful patterns from EEG signals and link them to biological changes associated with dementia. We aim to create models that are accurate, explainable, and trustworthy enough for clinical use. The work was carried out within the AI-MIND initiative (an EU Horizon 2020 project (Grant Agreement No. 964220), combining expertise from neuroscience, medicine, and artificial intelligence.
We analysed EEG recordings from several hospitals, together with clinical information such as age, genetic risk factors, and scores from memory tests. Using modern AI methods, such as transformer-based neural networks and self-supervised learning, we taught computers to recognise complex patterns in brain activity, including those invisible to the human eye. We also integrated EEG with clinical and genetic information to predict molecular biomarkers found in blood, such as phosphorylated tau (p-tau217), which are tightly linked to Alzheimer’s disease.
To support future clinical use, we made the models efficient and transparent. We applied model compression techniques so they can run on portable devices, and we developed DREAMS, an open-source tool that automatically generates standardized documentation (“model cards”) describing how each model works, what data it uses, and how reliable its predictions are.
This study shows that combining advanced AI with simple brain measurements can lead to non-invasive, accessible, and trustworthy tools for early dementia detection. Such technology could identify individuals at risk years before symptoms appear, enabling earlier intervention, personalized care, and an improved quality of life for millions worldwide.
