Discovering the neuronal dynamics in major depressive disorder using Hidden Markov Model

Wenhao Jiang, Shihang Ding, Cong Xu, Huihuang Ke, Hongjian Bo, Tiejun Zhao, Lin Ma, Haifeng LiFrontiers in Human Neuroscience 2023 · DOI

Abstract

In conclusion, the proposed HMM-MAR model has revealed its potential capability to capture the neuronal dynamics from EEG signals and interpret brain disease pathogenesis from the perspective of state transition. Compared with the previous machine-learning or deep-learning-based studies, which regarded the decoding model as a black box, this work has its superiority in the spatiotemporal pattern interpretability by utilizing the Hidden Markov Model.

Introduction

Major Depressive Disorder (MDD) is a leading cause of disability worldwide, yet its diagnosis remains largely clinical and subjective. While modern machine learning and deep learning models can classify MDD from EEG data with high accuracy, they often function as "black boxes," offering little insight into the underlying neurological mechanisms. This research, published in Frontiers in Human Neuroscience, presents a move away from these opaque diagnostic tools toward interpretable models of mental health. By applying Hidden Markov Models (HMM) to EEG data, researchers have identified specific patterns in how the brain transitions between states in patients with MDD, providing a window into the disease's pathogenesis.

Methods

The researchers employed a Hidden Markov Model with Multivariate Autoregressive observations (HMM-MAR) to decode EEG signals. This approach treats brain activity as a sequence of discrete, hidden states, assuming the brain transitions between a finite number of functional states (six, in this study) over time. Unlike standard HMMs, the Multivariate Autoregressive (MAR) component describes the actual signal dynamics—the "texture" of the EEG—within each state, capturing linear dependencies between 128 EEG channels.

Schematic workflow of the proposed HMM-MAR model.

The study utilized the MODMA dataset from Lanzhou University, comprising 128-channel scalp EEG data from 53 participants (24 MDD, 29 healthy controls). Participants performed a "dot-probe" task while viewing emotional face pairs (happy, sad, and fearful). Variational Bayesian inference was used to estimate model parameters, allowing for automatic feature extraction from the time-series data without manual engineering.

Key Results

The "Sadness" Signal and PHQ-9 Regression

The model successfully regressed PHQ-9 scores, a standard measure of depression severity. A critical finding was that brain dynamics in response to sad stimuli were the most predictive of depression severity. For the MDD group under sad stimuli, the model achieved a Root Mean Square Error (RMSE) of 4.04, significantly outperforming results for happy or fearful stimuli. This suggests that the "Markov property" of the brain—how the current state dictates the next—is most distinct in MDD patients when processing negative emotions.

Regression results for PHQ-9 scores across different emotional stimuli.

Neuronal "Inertia" and Connectivity

MDD patients exhibited significantly lower "switching rates" between brain states, suggesting a form of neuronal "inertia." Their neural dynamics appear more stable or "stuck" compared to healthy individuals, who transitioned between functional states more fluidly. Furthermore, MDD patients showed a reduced tendency to transition into active brain responses when presented with sad faces. Coherence analysis during sad stimuli also revealed a breakdown in connectivity between the frontal lobe and other regions, aligning with established theories of frontal lobe dysfunction in depression.

Discussion and Conclusion

By utilizing HMM-MAR, the researchers could visualize "connectivity maps" for each hidden state, showing exactly which brain regions were communicating during specific temporal windows. Power spectrum estimation revealed significant activation in the parietal and frontal lobes within the theta and alpha frequency bands during sadness-related states, providing a localized "fingerprint" of the depressive response.

While the study is limited by a relatively small sample size (n=53) and the need for high-density EEG equipment, it demonstrates the superiority of HMMs in providing spatiotemporal interpretability. Future work aims to map these hidden states to well-known functional networks like the Default Mode Network (DMN) and optimize the model for clinical integration.

Summary updated

Abstract

In conclusion, the proposed HMM-MAR model has revealed its potential capability to capture the neuronal dynamics from EEG signals and interpret brain disease pathogenesis from the perspective of state transition. Compared with the previous machine-learning or deep-learning-based studies, which regarded the decoding model as a black box, this work has its superiority in the spatiotemporal pattern interpretability by utilizing the Hidden Markov Model.
BibTeX
@inproceedings{jiang2023discovering,
  author = {Wenhao Jiang and Shihang Ding and Cong Xu and Huihuang Ke and Hongjian Bo and Tiejun Zhao and Lin Ma and Haifeng Li},
  title = {{Discovering the neuronal dynamics in major depressive disorder using Hidden Markov Model}},
  booktitle = {Frontiers in Human Neuroscience},
  year = {2023},
  doi = {10.3389/fnhum.2023.1197613},
}
Figure