A data-driven method to identify frequency boundaries in multichannel electrophysiology data

Background: Electrophysiological recordings of the brain often exhibit neural oscillations, defined as narrowband bumps that deviate from the background power spectrum. These narrowband dynamics are grouped into frequency ranges, and the study of how activities in these ranges are related to cognition and disease is a major part of the neuroscience corpus. Frequency ranges are nearly always defined according to integer boundaries, such as 4-8 Hz for the theta band and 8-12 Hz for the alpha band.

New method: A data-driven multivariate method is presented to identify empirical frequency boundaries based on clustering of spatiotemporal similarities across a range of frequencies. The method, termed gedBounds, identifies patterns in covariance matrices that maximally separate narrowband from broadband activity, and then identifies clusters in the correlation matrix of those spatial patterns over all frequencies, using the dbscan clustering algorithm. Those clusters are empirically derived frequency bands, from which boundaries can be extracted.

Results: gedBounds recovers ground truth results in simulated data with high accuracy. The method was tested on EEG resting-state data from Parkinson's patients and control, and several features of the frequency components differed between patients and controls.

Comparison with existing methods: The proposed method offers higher precision in defining subject-specific frequency boundaries compared to the current standard approach.

Conclusions: gedBounds can increase the precision and feature extraction of spectral dynamics in electrophysiology data.