Abstract
Non-suicidal self-injury (NSSI) is a critical indicator of adolescent mental health risks. While Electroencephalography (EEG) provides a non-invasive method for monitoring brain activity, its clinical application is hindered by label scarcity and high inter-individual variability. This paper introduces NSSI-Net, a novel semi-supervised Generative Adversarial Network (GAN) designed to detect NSSI behaviors. By integrating a spatiotemporal feature extractor with a multi-concept discriminator—accounting for signal quality, gender differences, and domain gaps—NSSI-Net achieves a state-of-the-art accuracy of 70.00% on a clinical dataset of 114 adolescent patients.
Introduction
Non-suicidal self-injury (NSSI) behaviors, such as cutting or burning without suicidal intent, are significant predictors of severe mental health outcomes in adolescents. EEG offers a high-resolution window into brain dynamics, yet two primary challenges persist: the high cost of expert-labeled clinical data and the inherent variability of brain signals across different demographics. Traditional machine learning models often fail to capture the complex spatiotemporal patterns necessary for robust diagnosis, especially when labeled data is limited. This work proposes a semi-supervised framework to leverage both labeled and unlabeled data while explicitly modeling demographic noise.
Methods
NSSI-Net addresses these challenges through a semi-supervised framework consisting of two main components: a spatiotemporal feature extractor and a multi-concept discriminator.
Spatiotemporal Feature Extraction
The model utilizes an encoder-decoder structure where a 2D-CNN captures spatial relationships between EEG channels, followed by a bi-directional Gated Recurrent Unit (BiGRU) to model temporal dynamics. Research findings indicated that this sequential extraction of spatial then temporal features outperforms parallel extraction, better mirroring the hierarchical structure of brain data.
Multi-Concept Discriminator
The defining feature of NSSI-Net is its multi-concept discriminator, which employs four specialized loss functions to refine the learned features:
- Signal-specific ($\mathcal{L}_{signal}$): Ensures high-quality, authentic feature generation.
- Gender-specific ($\mathcal{L}_{gender}$): Accounts for biological differences in EEG patterns between males and females.
- Domain-specific ($\mathcal{L}_{domain}$): Aligns features across labeled and unlabeled data to reduce domain gaps.
- Disease-specific ($\mathcal{L}_{disease}$): The final classifier distinguishing between patients with and without NSSI behaviors.
Results and Discussion
The model was evaluated on a self-collected EEG dataset of 114 adolescent patients diagnosed with Major Depressive Disorder (MDD). NSSI-Net achieved a mean accuracy of 70.00%, representing a 5.44% improvement over existing state-of-the-art semi-supervised methods like MixMatch (64.56%) and PARSE (61.77%).
Key Insights
- Demographic Modeling: The inclusion of a gender-specific discriminator was crucial. By explicitly modeling gender-based differences in brain activation, the network effectively ignored "gender noise" to focus on relevant disease signals.
- Semi-Supervised Efficiency: The model reached a performance plateau with only 75% of the source data labeled, suggesting that high diagnostic accuracy is achievable without exhaustive expert labeling.
Conclusion
NSSI-Net demonstrates the efficacy of multi-concept adversarial training in clinical EEG analysis. By effectively leveraging unlabeled data and modeling demographic variability, it provides a more accurate and robust tool for early NSSI detection. Future work will focus on addressing gender imbalances in clinical datasets and testing the model's generalizability in real-world clinical settings.
Summary updated