🧠 neuroscience

Canonical Hidden Markov Model Networks for Studying M/EEG

This paper introduces a publicly available, population-scale canonical Hidden Markov Model trained on 1849 MEG recordings to provide a common reference framework for analyzing diverse M/EEG datasets in both sensor and source spaces, thereby eliminating the need for computationally expensive, study-specific model training.

Original authors: Gohil, C., Huang, R., Higgins, C., van Es, M. W. J., Quinn, A. J., Vidaurre, D., Woolrich, M. W.

Published 2026-01-27

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Gohil, C., Huang, R., Higgins, C., van Es, M. W. J., Quinn, A. J., Vidaurre, D., Woolrich, M. W.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine the human brain as a bustling city with millions of people constantly moving, talking, and working together. Scientists use special cameras (called M/EEG) to watch this city from above, trying to understand how the neighborhoods (brain networks) switch on and off to handle different tasks like thinking, resting, or remembering.

For a long time, every time a scientist wanted to study a specific group of people, they had to build a brand new map from scratch. It was like every time you wanted to study traffic in a small town, you had to hire a team of cartographers to drive around for weeks, draw new roads, and create a unique map just for that one town. This was slow, expensive, and made it hard to compare the traffic in Town A with the traffic in Town B because the maps were drawn in different styles.

The Big Idea: A "Master Map"
This paper proposes a solution: instead of making a new map for every single study, let's create one universal "Master Map" (called a Canonical Hidden Markov Model).

Think of this Master Map as a standard set of LEGO instructions. Instead of every builder inventing their own unique way to snap bricks together, everyone uses the same official instruction book. This book describes the most common, reliable patterns of how the brain's "neighborhoods" switch on and off.

How They Built It
To create this Master Map, the researchers didn't just look at a few people. They gathered data from 1,849 different recordings (over 194 hours of brain activity!) from people aged 18 to 88. They watched these people both when they were resting and when they were doing tasks. By analyzing this huge crowd, they figured out the "standard" patterns of brain activity that happen across the general population.

How It Works in Practice
Now, if a scientist has a small, special dataset (a "boutique" study) with just a few patients, they don't need to build a new map. They can simply take their data and fit it onto the Master Map.

The "Source Space" Method: This is like looking at the city's internal street layout. To use the Master Map this way, the scientist must process their data exactly the same way the original team did (using the same "lenses" to see the brain's internal structure).
The "Sensor Space" Method: This is like looking at the city from a helicopter without needing to see the streets inside. The researchers showed you can use the Master Map directly on the raw camera data, skipping the complex step of mapping the internal streets. This makes it easier for more scientists to use the tool.

What They Did With It
The team tested this Master Map on three different small studies:

A group of people with Alzheimer's (resting state).
A group doing a working memory task.
A group using EEG (a different type of brain camera) while resting.

In all cases, the Master Map successfully described the brain activity, proving it works as a common language for different types of studies.

The Takeaway
The best part? The researchers have given this Master Map away for free. It is now an open-access resource. This means scientists everywhere can use the same standard set of brain networks to compare their patients with others, making research faster, cheaper, and much more consistent. Instead of everyone speaking a different dialect of brain science, they can all finally speak the same language.

Technical Summary: Canonical Hidden Markov Model Networks for Studying M/EEG

Problem Statement
Dynamic brain networks derived from magneto/encephalography (M/EEG) recordings offer critical insights into human brain activity. The Hidden Markov Model (HMM) is an established method for inferring reproducible, fast-switching brain networks across various cognitive and disease conditions. However, current research practices typically involve training bespoke HMMs on small, isolated datasets (often $N < 100$ ). This approach presents two primary limitations: it is computationally expensive to train a new model for every individual study, and analyzing datasets in isolation prevents direct comparison across different studies due to the lack of a common reference framework.

Methodology
To address these limitations, the authors propose a "canonical HMM" approach. Instead of training study-specific models, this method utilizes a single, pre-trained HMM that serves as a common reference for describing diverse datasets using the same set of networks.

The canonical HMMs were trained on a large-scale dataset comprising 1,849 MEG recordings (totaling 194 hours of data) from 621 participants aged 18–88 years. This training set captured population variability across both resting-state and task-based conditions. The models were developed for a range of model orders, specifically 4 to 16 states, and provided in two distinct domains:

Parcellated Source Space: Requires boutique datasets to undergo the same preprocessing and source reconstruction procedures as the training data.
Sensor Space: Requires the same sensor layout and preprocessing as the training data, allowing for the transfer of the parcel-level canonical HMM to sensor space without the need for individual source reconstruction.

Key Contributions and Results
The paper demonstrates the utility of this canonical approach through applications on three "boutique" datasets:

An Alzheimer's resting-state MEG dataset.
A working memory task MEG dataset.
A resting-state EEG dataset.

By applying the canonical HMM to these datasets, the authors show that different studies can be characterized using a unified set of networks. The study confirms that the canonical HMMs can be effectively transferred to sensor space, bypassing the computational and methodological complexities of source reconstruction for new studies, provided the sensor layout matches the training data.

Significance and Claims
The primary significance of this work lies in the creation of a standardized, open-access resource for the M/EEG community. The authors provide publicly available canonical HMMs for varying model orders, enabling researchers to compare individuals within and across a wide range of datasets using a consistent framework.

The paper claims that this approach offers a more efficient alternative to training new models for every study, reducing computational costs while potentially increasing the reliability of results by leveraging a model trained on a large, diverse population. The canonical HMMs serve as a common reference point, facilitating the description of different studies through the same set of brain networks, thereby fostering greater comparability in neuroimaging research.

Technical Summary: Canonical Hidden Markov Model Networks for Studying M/EEG

More like this