TL;DR
This paper introduces DepL-GCN, a novel deep learning model for classifying depression severity levels from EEG signals, addressing the challenges of subjective labeling and class imbalance. By incorporating sample confidence and minority sample penalty modules, DepL-GCN significantly improves multi-class depression level recognition accuracy, outperforming existing methods.
What Problem Does This Paper Address?
Major Depressive Disorder (MDD) is a significant global health issue, yet its diagnosis often relies on subjective clinical interviews and questionnaires. While EEG offers an objective, non-invasive method for neural monitoring, existing deep learning models primarily focus on binary classification (depressed vs. normal), lacking the granularity needed for assessing depression severity. This limitation stems from two key challenges: 1) the inherent subjectivity in depression-level labeling due to patient self-reports and potential biases, and 2) the significant class imbalance that arises when categorizing depression into multiple severity levels, where healthy individuals often vastly outnumber those in specific depression categories. This imbalance can lead models to favor majority classes, neglecting minority ones and hindering accurate severity assessment.
Methods
The core of this work is the DepL-GCN model, built upon the SSPA-GCN architecture, which is designed for EEG-based depression detection. DepL-GCN introduces two novel modules to tackle the aforementioned challenges:
- Sample Confidence Module: This module aims to mitigate the impact of subjective labeling. It dynamically assesses the confidence of each EEG sample's prediction during training. Samples with weak label alignment (i.e., low confidence) are progressively filtered out, reducing their influence on model training. This is achieved by calculating an "eL2 norm" (a measure of prediction error) and using it to derive a confidence value. The module's activation is strategically timed, starting after the model has learned simpler samples to avoid hindering the learning of genuinely difficult but correctly labeled data.
- Minority Sample Penalty Module: This module addresses the class imbalance issue. It increases the model's attention to samples belonging to minority classes (e.g., specific depression severity levels with fewer data points). This is done by assigning a penalty coefficient that is positively correlated with the prediction error (NeL2 norm) for misclassified minority samples. This penalty effectively upweights the loss for these samples, encouraging the model to learn them better.
The DepL-GCN model integrates these modules into the SSPA-GCN framework, optimizing a combined loss function that accounts for both class-level and domain-level losses, along with the contributions from the new modules.
Datasets
The study utilizes two publicly available EEG datasets:
- MODMA dataset: This dataset contains 128-channel resting-state EEG data from 53 clinically diagnosed depressive patients and 55 healthy controls. Subjects completed PHQ-9 and GAD-7 questionnaires.
- PRED+CT dataset: This dataset comprises EEG signals from depressive patients screened via the BDI and clinical interviews. Participants were categorized as healthy (BDI < 7) or patients (BDI >= 7).
The paper also discusses the classification criteria for depression levels based on PHQ-9 and BDI scores, defining categories such as Major, Moderate to Major, Moderate, Mild, and Normal.
Key Insights and Ideas
- Beyond Binary: The Need for Granularity: The paper strongly advocates for moving beyond simple binary depression detection to understanding severity levels, which is crucial for effective treatment and patient management.
- Subjectivity as a Data Problem: The authors frame the subjectivity in labeling not just as a human factor but as a data quality issue that can be addressed through algorithmic means like sample confidence filtering.
- Strategic Timing is Key: The effectiveness of the sample confidence module is highly dependent on when it's introduced during training. Introducing it too early can hinder learning, while too late can lead to overfitting.
- Prioritizing Minority Classes: The paper highlights that addressing class imbalance is not just about weighting, but about strategically prioritizing the learning of minority classes, especially when they are misclassified.
- Synergy of Modules: The combined effect of the sample confidence and minority sample penalty modules proved more effective than either module alone, demonstrating a synergistic improvement in performance.
Key Results
On the MODMA dataset, DepL-GCN achieved an accuracy of 81.13% for multi-class severity recognition. On the PRED+CT dataset, it achieved 81.36% accuracy. These results significantly outperformed the baseline SSPA-GCN model (71.70% on MODMA) and other state-of-the-art models, particularly in correctly classifying minority depression categories. Ablation studies confirmed the significant contributions of both the sample confidence module (improving accuracy to 79.25% on MODMA) and the minority sample penalty module (improving recall for minority classes).
Limitations and Future Work
The authors acknowledge several limitations and suggest future directions:
- Sample Confidence Calculation: The current method for calculating sample confidence relies on the eL2 norm. Future work could explore more sophisticated methods, potentially integrating more sample feature information for a more robust assessment.
- Clinical Application: While the model shows promising results, further research is needed to strengthen its application in real-world clinical environments, addressing practical challenges.
- Regression-Based Models: The paper briefly discusses regression-based models for depression level prediction, noting their potential for finer-grained assessment but also highlighting issues like asymmetric risk and the need for improved data quality and clinical interpretation. Future work could further explore and refine these regression approaches.
Summary updated