NSSI-Net: A Multi-Concept GAN for Non-Suicidal Self-Injury Detection Using High-Dimensional EEG in a Semi-Supervised Framework

TL;DR

This study tackles the automated detection of non-suicidal self-injury in adolescents with depression using high-dimensional EEG. It introduces NSSI-Net, a semi-supervised framework combining a spatial–temporal CNN–BiGRU encoder–decoder with a multi-concept discriminator that jointly addresses signal quality, gender, domain, and disease variations. On a self-collected cohort of 114 adolescents, the model achieves 70.0% accuracy under semi-supervised cross-subject validation and outperforms a range of traditional and deep-learning baselines, with ablation analyses confirming the distinct value of each discriminator. The work highlights neural patterns in NSSI, offers a pathway for early detection and intervention, and provides insights into gender- and domain-related EEG dynamics, while noting the need for larger and more balanced datasets to ensure broad applicability.

Abstract

Non-suicidal self-injury (NSSI) is a serious threat to the physical and mental health of adolescents, significantly increasing the risk of suicide and attracting widespread public concern. Electroencephalography (EEG), as an objective tool for identifying brain disorders, holds great promise. However, extracting meaningful and reliable features from high-dimensional EEG data, especially by integrating spatiotemporal brain dynamics into informative representations, remains a major challenge. In this study, we introduce an advanced semi-supervised adversarial network, NSSI-Net, to effectively model EEG features related to NSSI. NSSI-Net consists of two key modules: a spatial-temporal feature extraction module and a multi-concept discriminator. In the spatial-temporal feature extraction module, an integrated 2D convolutional neural network (2D-CNN) and a bi-directional Gated Recurrent Unit (BiGRU) are used to capture both spatial and temporal dynamics in EEG data. In the multi-concept discriminator, signal, gender, domain, and disease levels are fully explored to extract meaningful EEG features, considering individual, demographic, disease variations across a diverse population. Based on self-collected NSSI data (n=114), the model's effectiveness and reliability are demonstrated, with a 5.44% improvement in performance compared to existing machine learning and deep learning methods. This study advances the understanding and early diagnosis of NSSI in adolescents with depression, enabling timely intervention. The source code is available at https://github.com/Vesan-yws/NSSINet.

Figures (8)

• Figure 1: The proposed framework consists of two primary components: a spatial-temporal feature extraction module and a multi-concept discriminator. The feature extraction module leverages an encoder-decoder structure incorporating CNN and BiGRU to capture spatiotemporal features from EEG signals. The multi-concept discriminator involves four specific-designed discriminators: (1) Signal-specific discriminator distinguishes between real and generated EEG signals to enhance feature validity; (2) Gender-specific discriminator differentiates EEG patterns across genders to extract more generalized features; (3) Domain-specific discriminator aligns features from labeled source, unlabeled source, and target domains to reduce domain discrepancies; and (4) Disease-specific discriminator identifies EEG characteristics distinguishing adolescents with depression who engage in NSSI from those who do not.
• Figure 2: The sampling process strategy, considering the data balance and gender distribution in semi-supervised cross-subject cross-validation.
• Figure 3: The confusion matrices for different groups, where DN+ is considered as the positive class and DN- is considered as the negative class. (a) The overall confusion matrix. (b) The confusion matrix specific to the female subgroup. (c) The confusion matrix specific to the male subgroup. Each matrix displays the proportion of true positive (TP), true negative (TN), false positive (FP), and false negative (FN) predictions, with darker colors indicating higher percentages.
• Figure 4: The contribution of different model components (signal-specific, gender-specific, and domain-specific modules) towards the prediction accuracy of the model. Each vertical bar represents a specific combination strategy. Red, blue, green, and purple indicate the inclusion of signal-specific, gender-specific, domain-specific and disease-specific discriminator, respectively. Gray shows the absence. Combining more modules improves the model's accuracy, with the highest accuracy of 70.00% achieved when all components are included.
• Figure 5: The mean accuracy rates (%) of the model with different proportions of labeled source domain data.