Data-driven EEG subtyping of adolescents with non-suicidal self-injury: a machine-learning approach

Background

Non-suicidal self-injury (NSSI) is a clinically heterogeneous condition frequently conceptualized through functional models distinguishing intrapersonal (e.g., affect regulation and self-punishment) and social (e.g., interpersonal influence or communication) reinforcement processes. Although these models provide a useful psychological framework, biologically informed stratification of NSSI remains limited. Identifying neurophysiological subtypes may improve understanding of underlying mechanisms and inform more targeted interventions. Resting-state electroencephalography (EEG) offers a noninvasive and cost-effective method to examine intrinsic brain activity and neural network organization. Data-driven analytic approaches applied to EEG data may help uncover latent subgroups that are not readily detectable through symptom-based classification alone. Therefore, we applied unsupervised machine-learning techniques to resting-state EEG data to derive neurophysiological subgroups and characterize their associated clinical, personality, and cognitive profiles.

Methods

This retrospective study included 104 female adolescents and young adults (9–24 years) hospitalized for NSSI-related evaluation between 2018 and 2023. Resting-state EEG was recorded under eyes-closed conditions and preprocessed to extract quantitative spectral features representing neural activity across frequency bands. These features were analyzed using a two-step unsupervised framework. First, a Self-Organizing Map (SOM) was employed to visualize high-dimensional EEG feature patterns within a topologically structured representation, facilitating exploration of intrinsic structure in the data. Second, K-means clustering was applied to partition participants according to similarity in EEG features. The optimal cluster number was determined using the elbow method based on within-cluster sum of squares (WCSS). To enhance robustness against stochastic initialization effects inherent to K-means, clustering was repeated 1000 times and average WCSS values were examined. A clear inflection point at k=3 indicated the most stable and parsimonious solution. Between-group differences in multidomain clinical, personality, and cognitive domains were evaluated using multivariate analysis of covariance (MANCOVA), adjusting for antidepressant and antipsychotic use before and during hospitalization.

Results

Three EEG-defined subgroups were identified: Group 0 (Intrapersonal functions; n=37), Group 1 (Social functions; n=30), and Group 2 (Low Symptom Profile; n=37). Significant between-group differences were observed in depressive and anxiety-related measures. BDI scores differed across groups (F=8.27, p<0.001, partial η²=0.32), with Groups 0 and 1 scoring higher than Group 2. Similar patterns were observed for BAI (F=7.38, p<0.001, partial η²=0.33) and ASI (F=12.54, p<0.001, partial η²=0.46). Across CBCL 6–18 domains, significant effects were detected (F=3.22–5.74, p=0.01–0.05). Group 0 demonstrated broad elevations in Total Problems, Anxious/Depressed, Anxiety Problems, Obsessive–Compulsive Symptoms, PTSD Problems, Thought Problems, and related domains compared to Group 2 (all post-hoc p≤0.04). In contrast, Group 1 showed more selective elevations relative to Group 2, including phobic anxiety (SCL-90 PHOB, F=3.30, p=0.05), obsessive content (MMPI A-obs, F=4.24, p=0.02), and temperament subscales (JTCI NS2 and RD2, both p<0.05). Cognitive measures also differed (Digit Span/Stroop total, F=4.10, p=0.02, partial η²=0.16), with Group 2 demonstrating better attentional performance than Group 1.

Conclusions

Unsupervised EEG-based clustering identified three reproducible neurophysiological subtypes corresponding to distinct clinical profiles aligned with functional reinforcement models. These findings highlight biologically informed heterogeneity in NSSI and suggest that EEG-based stratification may enhance precision-oriented assessment and support more targeted, mechanism-informed intervention strategies.

Keywords: non-suicidal self-injury, adolescents, resting-state EEG, machine learning, unsupervised clustering

References

No references