Deep-Asymmetry: Asymmetry Matrix Image for Deep Learning Method in Pre-Screening Depression

Min Kang, Hyunjin Kwon, Jinhyeok Park, Seokhwan Kang, Youngho LeeSensors 2020 · DOI

TL;DR

This paper introduces "Deep-Asymmetry," a novel deep learning approach for pre-screening depression by transforming EEG asymmetry into a matrix image. This method achieves a remarkable 98.85% accuracy, outperforming previous techniques and offering a more objective and accessible diagnostic tool.

What Problem Does This Paper Address?

Diagnosing Major Depressive Disorder (MDD) is crucial, but current methods heavily rely on subjective consultations with psychologists or psychiatrists. This subjectivity can lead to misdiagnosis, especially with "masked depression" where symptoms don't fit typical presentations. Furthermore, the stigma surrounding mental illness often deters individuals from seeking help. While objective physiological markers are sought, existing EEG-based machine learning approaches often require complex, manual feature extraction from raw, one-dimensional data. This paper aims to bridge this gap by developing an automated, objective, and highly accurate method for depression pre-screening using EEG.

Methods

The core innovation of this study is the "Deep-Asymmetry" methodology. Instead of directly processing raw EEG data, the researchers convert EEG asymmetry into a 2D matrix image. This asymmetry is calculated based on the relative power differences between EEG channels across various frequency bands (delta, theta, alpha, beta). Specifically, the relative power of each channel is computed, and then the difference between pairs of channels is quantified using a specific formula. This difference is then visualized as a matrix where each cell's color represents the degree of asymmetry between two channels.

This asymmetry matrix image serves as the input for a Convolutional Neural Network (CNN). The CNN architecture consists of three layers: a 2D convolution layer, a pooling layer, and a fully connected layer. Key components like ReLU activation, batch normalization (to prevent gradient vanishing), and dropout (to prevent overfitting) are employed. The model is trained using the Adam optimizer with a binary cross-entropy loss function.

Datasets

The study utilized an open-access EEG dataset collected at the Hospital Universiti Sains Malaysia (HUSM). This dataset comprises EEG information from 34 depressed patients (diagnosed with DSM-IV) and 30 healthy controls. The data was collected during eye-closed (EC) and eye-opened (EO) states. The dataset was segmented into 4-second epochs for analysis.

Key Insights and Ideas

  • EEG Asymmetry as a Biomarker: The study reinforces the notion that brain asymmetry, particularly in the alpha band, is a significant biomarker for depression. This insight drives the core of their methodology.
  • Image Representation for CNNs: Transforming complex time-series EEG data into a visually interpretable 2D image (the asymmetry matrix) allows for the effective application of powerful CNN architectures, which excel at spatial feature extraction.
  • Automated Feature Extraction: The "Deep-Asymmetry" method automates the feature extraction process, moving away from manual, potentially error-prone, feature engineering. This makes the diagnostic tool more accessible and less reliant on expert knowledge.
  • Alpha Band Dominance: The alpha frequency band consistently yielded the highest performance metrics, suggesting its particular relevance in capturing depression-related EEG patterns through asymmetry.

Key Results

The proposed "Deep-Asymmetry" method demonstrated exceptional performance:

  • Eye-Closed (EC) Dataset:
    • Alpha Band: Achieved an accuracy of 98.85%, sensitivity of 99.15%, and specificity of 98.51%. This was the best-performing band.
  • Overall, the alpha band-based asymmetry image achieved the highest accuracy among all tested bands and datasets.
  • Eye-Opened (EO) Dataset:
    • Alpha Band: Achieved an accuracy of 97.56%, sensitivity of 98.75%, and specificity of 96.31%.
  • ROC Curves: The Receiver Operating Characteristic (ROC) curves further supported the effectiveness of the alpha band, showing strong classification performance with an average Area Under the Curve (AUC) of 0.9879 in the EC dataset.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Dataset Size: The study used a relatively small dataset, which they note is a limitation and could lead to overfitting. Generalization to a larger population requires further evaluation with more extensive datasets.
  • Generalization: The pilot nature of the study means that broader validation is needed to confirm the model's generalization capabilities.
  • Correlation with Clinical Scores: Future work could explore the correlation of the classification results with clinical depression scores, such as the Beck Depression Inventory (BDI).
  • Model Overfitting: While techniques like batch normalization and dropout were used, further research could focus on more robust methods to overcome potential overfitting.
  • Expanding Data Modalities: Investigating the use of other types of EEG data or combining EEG with other physiological signals could enhance the model's performance.
  • Real-world Deployment: The authors envision this method as a tool for nonprofessionals, suggesting future work could focus on making it even more user-friendly for pre-screening purposes.

Summary updated

Abstract

To have an objective depression diagnosis, numerous studies based on machine learning and deep learning using electroencephalogram (EEG) have been conducted. Most studies depend on one-dimensional raw data and required fine feature extraction. To solve this problem, in the EEG visualization research field, short-time Fourier transform (STFT), wavelet, and coherence commonly used as method s for transferring EEG data to 2D images. However, we devised a new way from the concept that EEG's asymmetry was considered one of the major biomarkers of depression. This study proposes a deep-asymmetry methodology that converts the EEG's asymmetry feature into a matrix image and uses it as input to a convolutional neural network. The asymmetry matrix image in the alpha band achieved 98.85% accuracy and outperformed most of the methods presented in previous studies. This study indicates that the proposed method can be an effective tool for pre-screening major depressive disorder patients.
BibTeX
@inproceedings{kang2020deep,
  author = {Min Kang and Hyunjin Kwon and Jinhyeok Park and Seokhwan Kang and Youngho Lee},
  title = {{Deep-Asymmetry: Asymmetry Matrix Image for Deep Learning Method in Pre-Screening Depression}},
  booktitle = {Sensors},
  year = {2020},
  doi = {10.3390/s20226526},
}
Figure