Reliability of resting-state EEG spectral power - advantage of normalization is not guaranteed

As a noninvasive and handy technique, EEG is widely used to study brain functions. However, while spectral analysis in terms of band power is extensively applied, the reliability of extracted features was seldom confirmed. The current study investigates the reliability of EEG spectral features in a resting-state setting. Resting state EEG is task-free and easy to collect, making it feasible for long-term monitoring of neurodegenerative diseases in clinical situations, for which reliability is a crucial prerequisite. Linear mixedeffect model (LMM) was employed to analyze eyes-open and eyes-closed resting-state data (three minutes each), collected in 16 elderly subjects (gender-balanced) across two sessions separated by a month on average. Absolute and relative band powers were extracted, with the relative band power being computed by normalizing the absolute power with respect to the total band power (1-45 Hz). Their reliability over the two sessions was evaluated by intraclass correlation (ICC). The major advantage of LMM is that subject-specific components such as baseline activities can be extracted. Results showed that despite being a normalized feature, the relative power did not exhibit its advantages consistently across all spectral bands. For example, absolute power in alpha band attained higher reliability than relative power over the whole scalp, but the pattern was reversed in beta band. While huge enhancements were found by using relative power in eyes-open beta band, the normalization greatly decreased the reliability in eyes-open theta band. Such inconsistency raised the need of careful consideration in feature selection based on research aim.