Dementia is a progressive neurodegenerative condition for which early detection remains a major clinical challenge. Growing evidence suggests that functional brain alterations precede clinical symptoms by several years, making functional MRI a promising modality for identifying individuals at risk before cognitive decline becomes apparent.
The primary research question of this project is whether latent representations of brain activity learned from healthy participants can be used to better detect early functional signatures associated with future dementia. Specifically, we aim to investigate whether deviations from normative functional brain patterns, captured through a foundation model trained exclusively on healthy individuals, are predictive of later dementia diagnosis.
The main objectives of the study are: (i) to derive low-dimensional brain embeddings from rs-fMRI data using Karavela’s foundation model trained on healthy participants; (ii) to characterise how these embeddings differ in individuals who later develop dementia compared with those who remain cognitively healthy; and (iii) to assess the predictive value of these representations for identifying individuals at increased risk of dementia years before diagnosis.
The scientific rationale is grounded in recent work demonstrating on the UK biobank dataset that early disruptions in large-scale brain networks, particularly within the default mode network, can be detected using resting-state fMRI and are associated with future dementia (e.g. Ereira et al., 2024).
This research will contribute to a better understanding of early functional brain changes associated with dementia and may support the development of scalable, data-driven biomarkers for early risk stratification in large population cohorts such as UK Biobank.