State-space Modeling of Neural Oscillations: Toward Assessing Alzheimer’s Disease Neuropathology with Sleep EEG
The recent development of Alzheimer's disease (AD) modifying therapies has led to a heightened need for early diagnostic methods. Studies over the past two decades have highlighted brain waves during sleep as a window of opportunity to assess neural activity impacted by AD neuropathology before...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Published: |
Massachusetts Institute of Technology
2024
|
| Online Access: | https://hdl.handle.net/1721.1/153679 https://orcid.org/0000-0002-6688-8693 |
| Summary: | The recent development of Alzheimer's disease (AD) modifying therapies has led to a heightened need for early diagnostic methods. Studies over the past two decades have highlighted brain waves during sleep as a window of opportunity to assess neural activity impacted by AD neuropathology before clinical symptoms emerge. However, many existing practices in sleep electroencephalography (EEG) analysis are based on visual inspection and the ubiquitous Fourier transform. These approaches face several methodological challenges demonstrated in this thesis, and they have become increasingly limiting as the scientific questions become more complicated, as is the case with studying sleep oscillations in preclinical AD patients.
In this thesis, I developed a set of methods centered around a parametric oscillator model. These new algorithms build on previous ideas of state-space modeling of neural oscillations and enable novel solutions to some fundamental questions in sleep EEG, which are crucial to address for its interpretability in older adults. First, a previous data-driven greedy search method is improved to identify the number of oscillations and more reliably characterize them through time-domain modeling. I applied this updated approach to distinguish between slow and delta oscillations during sleep and observed clearer associations of delta oscillations. Second, a new switching state-space solution extends the application of oscillator modeling to time varying neural oscillations. I derived a probabilistic detection of sleep spindles that automatically adjusts to individualized spindle characteristics. This method can work on data as short as a few seconds and extracts complex spindle activity, providing highly stable spindle property estimates. Third, a semi-automated pipeline is developed to build realistic four-layer forward models including the cerebrospinal fluid in each individual to account for cortical atrophy. A simulation study showed that cortical atrophy alone produces modest reductions in scalp EEG ~2dB while also spreading the source currents, suggesting that individualized head models might need to be employed to study sleep EEG. To this end, a general dynamic source localization solution is derived, with full capacities to accommodate spatial patterns in the cortex while maintaining efficient parameter estimation. A proof of concept analysis showed successful localization of multiple simultaneous oscillations in real EEG recordings. This work advances our abilities to analyze sleep EEG data from first principles of statistical modeling and signal processing, paving ways for an individualized and non-invasive assessment of sleep oscillations in older adults. These methods may provide a critical step toward rigorously inferring the underlying neural activity that is likely altered by early AD neuropathology. |
|---|