Multi-scale convolutional recurrent neural network for psychiatric disorder identification in resting-state EEG

Weizheng Yan, Linzhen Yu, Dandan Liu, Jing Sui, Vince D. Calhoun, Zheng LinFrontiers in Psychiatry 2023 · DOI

Abstract: The MsRNN demonstrated a capability in extracting discriminative rsEEG biomarkers for psychiatric disorder classification, indicating its potential to facilitate our understanding of psychiatric disorders and monitoring interventions.

Introduction

Psychiatric diagnosis currently relies almost entirely on phenomenology—the observation of clinical symptoms and patient self-reporting. Because disorders like Bipolar Disorder (BP) and Major Depressive Disorder (MDD) often share overlapping symptoms at onset, misdiagnosis is common; roughly 60% of BP patients are initially diagnosed with MDD.

While neuroimaging tools like fMRI offer biological insights, they are expensive and difficult to implement in routine clinical settings. Resting-state EEG (rsEEG) is a much more affordable and accessible alternative, but traditional analysis often struggles to capture the complex, non-linear spatiotemporal patterns—how brain activity changes across both different regions and over time—that distinguish one disorder from another.

Methods

The researchers proposed a Multi-scale Convolutional Recurrent Neural Network (MCRNN) architecture designed to handle the raw complexity of EEG time series data without needing manual feature engineering.

The overall pipeline of the study, illustrating the flow from dataset collection and preprocessing to model training and clinical output.

The study utilized a database of 327 individuals, including patients with Schizophrenia (SZ, n=70), Bipolar Disorder (BP, n=72), Major Depressive Disorder (MDD, n=138), and Healthy Controls (NC, n=47). The data consisted of 5 minutes of 64-channel rsEEG recorded while participants were awake with their eyes closed.

The MCRNN model consists of two key stages:

  1. Multi-scale 1D Convolutional Layers: Instead of using a single filter size, the model uses three different scales of filters (lengths of 2, 4, and 8) to extract spatial features from the 64 EEG channels. This allows the model to capture brain activity patterns occurring at different spatial resolutions.
  2. Gated Recurrent Units (GRU): After the spatial features are extracted and pooled, they are fed into a GRU module. GRUs are a type of recurrent neural network excellent at processing sequential data, allowing the model to "remember" the temporal flow of brain signals.
Detailed architecture of the MCRNN, showing the multi-scale spatial filters and the temporal processing GRU units.

To ensure interpretability, the researchers used Uniform Manifold Approximation and Projection (UMAP) to map high-dimensional features into a 2D "psychiatric spectrum" and employed an occlusion strategy to identify the most discriminative brain regions.

Key Results and Insights

The MCRNN achieved high performance in binary classification tasks:

  • SZ vs. NC: 91.3% accuracy
  • MDD vs. NC: 84.9% accuracy
  • BP vs. NC: 78.6% accuracy

In a four-class classification (distinguishing all groups simultaneously), the model achieved 68.2% accuracy, a significant result given the high comorbidity of these conditions.

The UMAP visualization revealed a biological "severity spectrum," where healthy controls clustered at one end, followed by MDD and BP, with Schizophrenia at the far end. This suggests these disorders exist on a neurobiological continuum rather than as isolated categories. Anatomical analysis identified the most important biomarkers in the occipital lobe (O1, O2), parietal lobe (P3, P7, P8), and frontal lobe (Fp1, F8), regions associated with cognitive and sensory processing disruptions.

Conclusion and Future Work

The MCRNN demonstrated that internal feature representations in Schizophrenia patients correlated with clinical PANSS scores, indicating the model captures meaningful markers of disease progression. While the current sample size is a solid foundation, future work will focus on larger, more diverse datasets and the inclusion of other disorders like ADHD or anxiety. Integrating demographic data and longitudinal monitoring could further transform this tool into a robust system for tracking patient response to treatment over time.

Summary updated

Abstract

The MsRNN demonstrated a capability in extracting discriminative rsEEG biomarkers for psychiatric disorder classification, indicating its potential to facilitate our understanding of psychiatric disorders and monitoring interventions.
BibTeX
@inproceedings{yan2023multi,
  author = {Weizheng Yan and Linzhen Yu and Dandan Liu and Jing Sui and Vince D. Calhoun and Zheng Lin},
  title = {{Multi-scale convolutional recurrent neural network for psychiatric disorder identification in resting-state EEG}},
  booktitle = {Frontiers in Psychiatry},
  year = {2023},
  doi = {10.3389/fpsyt.2023.1202049},
}
Figure