TL;DR
This paper introduces a novel deep learning framework, MRCNN-RSE, that accurately identifies depressive disorder (DD) using only six frontal electroencephalogram (EEG) channels. This approach significantly improves the practicality of DD diagnosis by reducing data acquisition time and complexity compared to high-density EEG methods.
What Problem Does This Paper Address?
Diagnosing depressive disorder (DD) is currently a subjective process heavily reliant on clinical psychiatrists' experience and self-reported scales. This approach lacks objectivity, accuracy, and practicality, leading to potential misdiagnoses and delayed treatment. While neuroimaging techniques like EEG show promise for objective diagnosis, high-density EEG systems, often used in research, are cumbersome and time-consuming, hindering widespread clinical application. This paper aims to bridge this gap by developing an accurate and practical DD diagnostic method using a simplified EEG setup (six frontal channels) combined with advanced deep learning.
Methods
The core of this research lies in developing and evaluating two deep learning models for DD recognition: MRCNN-LSTM and MRCNN-RSE. Both models leverage a Multi-Resolution Convolutional Neural Network (MRCNN) to extract time-frequency features from EEG signals.
- MRCNN: This component uses parallel convolutional branches with varying kernel sizes to capture features at different scales. It incorporates Batch Normalization and ReLU activation for feature enhancement.
- MRCNN-LSTM: This model further processes the MRCNN features using a Long Short-Term Memory (LSTM) network to capture temporal dependencies in the EEG data.
- MRCNN-RSE: This model enhances the MRCNN features with a Residual Squeeze-and-Excitation (RSE) block. The RSE block adaptively recalibrates feature responses by modeling interdependencies between features, aiming to highlight the most discriminative ones. This is achieved through convolutional layers, adaptive pooling, and fully connected layers with specific activation functions.
The study also investigated the impact of different EEG frequency bands (theta, alpha1, alpha2, beta, and combined bands) on classification performance.
Datasets
The study utilized EEG data collected from:
- 41 patients diagnosed with depressive disorder (DD).
- 34 healthy controls (HCs).
All participants underwent the Hamilton Depression Scale (HAMD) assessment. The data consisted of 10-minute resting-state EEG recordings. No new benchmark datasets were introduced; the study used its collected dataset for evaluation.
Key Insights and Ideas
- Frontal EEG Sufficiency: The research demonstrates that a limited set of six frontal EEG channels can achieve high accuracy in DD recognition, challenging the notion that high-density EEG is always necessary for effective diagnosis. This significantly boosts the practicality of EEG-based DD screening.
- Beta Rhythm's Importance: The study highlights that the beta rhythm (13-30 Hz) shows significantly altered activity in DD patients, leading to higher classification accuracy compared to other rhythms. This reinforces the role of beta rhythm in DD pathogenesis and diagnosis.
- MRCNN-RSE Superiority: The MRCNN-RSE model consistently outperformed other tested deep learning models, including MRCNN-LSTM, EEGNet, DeprNet, 1DCNN-LSTM, and 2DCNN-LSTM, achieving the highest accuracy. This suggests the RSE block's effectiveness in adaptively selecting discriminative features is crucial for DD detection.
- High-Frequency Bands Advantage: Utilizing higher frequency bands (e.g., 8-30 Hz) for DD diagnosis not only yielded better classification performance but also potentially reduced EEG signal preprocessing time and improved the signal-to-noise ratio.
Key Results
- The proposed MRCNN-RSE model achieved a remarkable 98.48 ± 0.22% accuracy for DD recognition using 8-30 Hz EEG signals.
- The beta rhythm (13-30 Hz) showed the highest accuracy among individual rhythms, with MRCNN-RSE achieving 97.13 ± 0.49% accuracy.
- The MRCNN-RSE model outperformed other deep learning models, including MRCNN-LSTM (95.38% accuracy), EEGNet (90.07% accuracy), and others, on the 8-30 Hz EEG dataset.
- The study found that the 8-30 Hz frequency band provided superior accuracy compared to the broader 4-30 Hz band, indicating the benefit of focusing on higher frequencies.
Limitations and Future Work
- Limited Sample Size: The authors acknowledge that the study included a limited number of participants (41 DD patients and 34 HCs). This limitation might affect the generalization performance of the MRCNN-RSE model.
- Future Work: The researchers plan to collect more samples in future studies to further validate and promote the practical implementation of their classification approach. They also aim to continue exploring the effectiveness of different EEG frequency bands and deep learning architectures for DD diagnosis.
Summary updated