Abstract

Understanding sleep stages and their underlying cognitive transitions is essential for advancing our knowledge of brain function. While sleep staging has traditionally relied on electroencephalogram (EEG), this study explores the feasibility of classifying sleep stages using only functional Magnetic Resonance Imaging (fMRI) data. We investigate the use of Hidden Markov Models (HMM) and Hierarchical Dirichlet Process Hidden Markov Models (HDP-HMM) to capture temporal brain-state dynamics from task-free fMRI recordings. A series of pipelines—including sliding window, two-stage majority filtering, and prototype-based mapping via the Hungarian algorithm—were implemented to align model-inferred states with EEG-defined sleep stages. Among the twelve tested pipelines, the HMM with majority filtering and prototype mapping achieved the highest overall classification accuracy while maintaining interpretability and computational efficiency. Analyses of transition probabilities, dwell times, and fractional occupancy revealed that the HMM-derived states exhibited weak correspondence with EEG sleep stages in both temporal structure and functional connectivity patterns. Limitations emerged due to the low temporal resolution and high dimensionality of fMRI data, which hindered the detection of short-lived or rare sleep stages such as REM. These results demonstrate the potential of unsupervised HMM-based approaches for fMRI-only sleep staging and highlight the need for future enhancements through semi-Markov models, adaptive smoothing, and integration of dimensionality reduction techniques to better capture transient cognitive states.

Committee Chair

Soumendra Lahiri

Committee Members

Muriah D Wheelock, Debashis Mondal

Degree

Master of Arts (AM/MA)

Author's Department

Statistics

Author's School

Graduate School of Arts and Sciences

Document Type

Thesis

Date of Award

5-2025

Language

English (en)

Download

Share

COinS