Abstract

The prevention of neurodegenerative diseases is an emerging field that relies on understanding the evolution of their symptoms such as cognitive impairment. Although there are no standard indicators for detecting the early signs of dementia, it is well-established that several memory-related processes occur during sleep. Therefore, recent works have studied sleep through brain activity, focusing on the sleep stages defined by the American Academy of Sleep Medicine (AASM). In this thesis, we demonstrate that relying on AASM stages for detecting dementia provides suboptimal results and propose a new and more informative set of sleep stages discovered naturally by a Gaussian Mixture (GM) model operating on electroencephalogram (EEG) features. The improved performance of our approach emerges from simple logistic regression classifiers trained on EEG ‑derived descriptors from healthy subjects, as well as patients with mild cognitive impairment and severe dementia, highlighting the richer information content captured by our GM‑based stages. Moreover, we show empirically that differences between levels of dementia reside primarily in patterns of brain activity rather than in the time spent in each sleep stage. These findings indicate that more expressive characterizations of sleep exist beyond the conventional AASM framework. In future work, given the promising results from this thesis, it would be interesting to study other pathologies, as well as other techniques to extract information from sleep.