Canonical Hidden Markov Model Networks for Studying M/EEG

Dynamic brain networks identified in magneto/encephalography (M/EEG) recordings provide new insights into human brain activity. One established method uses Hidden Markov Models (HMMs) and has been shown to infer reproducible, fast-switching brain networks in a variety of cognitive and disease conditions. Often these studies are done on small bespoke (boutique) datasets (N<100) and analysed in isolation of other M/EEG datasets. Instead of training a new model for each boutique study, which is computationally expensive, we propose the use of a canonical HMM. This provides a common reference through which different studies can be described using the same set of networks. We provide HMMs for a range of model orders (4-16 states) in parcellated source space and sensor space. These HMMs were trained on 1849 MEG recordings (N=621, 18-88 years old, 194 hours), capturing population variability in both rest and task data. We illustrate applications of this canonical HMM approach in parcellated source space using boutique MEG and EEG datasets. Applying the canonical HMM in parcel space requires the boutique dataset to be preprocessed and source reconstructed in the same way as the canonical HMM training data. Apply the canonical HMM in the sensor space requires the same sensor layout and preprocessing as the canonical HMM training data. The canonical HMMs have been made publicly available as an open-access resource, providing sets of canonical brain networks that can be used to compare individuals within and across a range of datasets.