Stable EEG Spatiospectral Patterns Estimated in Individuals by Group Information Guided NMF

• Published in: Brain topography Vol. 38; no. 3; p. 36
• Main Authors: Zhou, Tianyi, Li, Xuan, Wang, Juan, Li, Zheng, Yin, Liyong, Yin, Bowen, Geng, Xinling, Li, Xiaoli
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.05.2025; Springer Nature B.V
• Subjects: Algorithms; Alzheimer Disease - physiopathology; Alzheimer's disease; Biomedical and Life Sciences; Biomedicine; Brain - physiology; Brain - physiopathology; Brain Mapping - methods; Cognitive ability; EEG; Electroencephalography; Electroencephalography - methods; Female; Humans; Male; Neurodegenerative diseases; Neurology; Neurosciences; Original Paper; Oscillations; Psychiatry; Signal Processing, Computer-Assisted; Nonnegative Matrix Factorization; EEG Patterns; EEG; Group-information Guided NMF
• ISSN: 0896-0267; 1573-6792; 1573-6792
• DOI: 10.1007/s10548-025-01110-5

Electroencephalographic (EEG) oscillations occur across a wide range of spatial and spectral scales, and analysis of neural rhythmic variability have attracted recent attention as markers of development, intelligence, cognitive states and neural disorders. Nonnegative matrix factorization (NMF) has been successfully applied to multi-subject electroencephalography (EEG) spectral analysis. However, existing group NMF methods have not explicitly optimized the individual-level EEG components derived from group-level components. To preserve EEG characteristics at the individual level while establishing correspondence of patterns across participants, we present a novel framework for obtaining subject-specific EEG components, which we term group-information guided NMF (GIGNMF). In this framework, group information captured by standard NMF at the group level is utilized as guidance to compute individual subject-specific components through a multi-objective optimization strategy. Specifically, we propose a three-stage framework: first, group-level consensus EEG patterns are derived using standard group NMF tools; second, an optimal procedure is implemented to determine the number of components; and finally, the group-level EEG patterns serve as references in a new one-unit NMF employing a multi-objective optimization solver. We test the performance of the algorithm on both synthetic signals and real EEG recordings obtained from Alzheimer’s disease data. Our results highlight the feasibility of using GIGNMF to identify EEG spatiotemporal patterns and present novel individual electrophysiological characteristics that enhance our understanding of cognitive function and contribute to clinical neuropathological diagnosis.