Javier Garcia Ciudad
Abstract:
The purpose of this project is to expand our knowledge about the electrophysiological features of sleep, with a particular focus on establishing links and differences between human and mouse sleep in both healthy and narcoleptic phenotypes. Narcolepsy is a sleep disorder characterized by excessive daytime sleepiness. Mouse models are often used to study narcolepsy by introducing specific pathological changes with gene manipulation techniques. Both in humans and mice, sleep and narcolepsy are often studied using electrophysiological signals. Still today, these signals are mainly analyzed by manual annotation of different sleep stages. In recent years, deep learning scoring models have been introduced, though without becoming widely implemented.
These models apply just to humans or just to mice, which is partly motivated by a lack of understanding of how much human and mouse sleep have in common. Finding similarities between both would support the development of common scoring models. More importantly, it would allow causal links to be made between the specific pathological changes modeled in mice and the human disease, which is one of the major challenges in narcolepsy research. In addition, finding electrophysiological signatures of narcolepsy or other factors such as age or gender would enhance our understanding of narcolepsy and sleep.
For this purpose, sleep signals will be studied using state-of-the-art deep learning methods. Sleep scoring models based on transformers and convolutional and recurrent neural networks will be studied to investigate how well they translate between the human and mouse domain. In addition, representation learning using variational autoencoders and contrastive learning techniques will be employed to learn compact representations of sleep signals, with the goal of providing species-invariant representations and identifying individual variabilities from the signals. The learned representations will be projected to lower- dimensional latent spaces, in which evaluating the distance between groups. Finally, explainable AI techniques will be investigated to extract insights from the models used, which could reveal EEG biomarkers of species, disease state and other individual variabilities.