Social Pain Disrupts Emotion-Action Brain-State Dynamics in Adolescents with Non-Suicidal Self-Injury

Introduction

• NSSI is a major concern among adolescents with depression and is associated with greater clinical severity and prospective suicidal-behavior risk [3, 4], while identification relies largely on self-report and interviews [5].Epidemiological studies estimate a 17.7% lifetime prevalence in the general adolescent population, with substantially higher rates among adolescents with depression [1, 2].Source: Introduction · p. 1
• Prior work has focused mainly on regional activation or static measures, leaving unclear how brain states transition as social distress unfolds in adolescents with NSSI.Social rejection and physical pain engage partially overlapping neural systems [12], but the dynamics linking social distress to emotional, interoceptive, and action-related processes remain insufficiently characterized.Source: Introduction · p. 2
• EEG microstates capture rapidly changing whole-brain functional configurations, but conventional aggregate measures may miss higher-order dependencies among evolving states.Conventional summaries include duration, occurrence, coverage, and pairwise transition probabilities; sequence-based modeling can directly represent higher-order temporal dependencies [22].Source: Introduction · p. 2
• The study combines experimental pain conditions, EEG microstate analysis, and interpretable deep sequence modeling to identify NSSI-related dynamics and characterize their functional and behavioral relevance.It examines social pain, physical pain, and rest, alongside source reconstruction, functional decoding, time-resolved analysis, and mediation analysis.Source: Introduction · p. 2

Results

• The analysis modeled EEG microstate sequences across social pain, physical pain, and rest, using source reconstruction, time-resolved analysis, and mediation analysis to characterize the identified dynamics.The proposed framework uses microstate sequence modeling to capture higher-order transition dependencies.Source: Results · pp. 2–3, Figure 1 · p. 3

Disrupted MS3-MS5 transitions characterize social pain processing in NSSI

• During social pain, DN+ adolescents showed weaker bidirectional MS3↔MS5 transitions than DN− adolescents, while sex differences were limited to MS6→MS7.The MS3↔MS5 finding converged with conventional microstate analyses.Source: Disrupted MS3-MS5 transitions… · p. 4

MS3-MS5 dynamics link emotionalinteroceptive processing to action preparation

• MS3 was associated with emotional/interoceptive processing and MS5 with action preparation, linking their reduced transitions to disrupted emotion-action coupling.The reported profiles implicate cingulate and medial frontal regions for MS3 and precentral and adjacent premotor regions for MS5.Source: MS3-MS5 dynamics link… · p. 4

Time-resolved dynamics reveal an early action-preparation bias followed by late emotional recruitment

• DN+ adolescents showed greater action-preparation-state recruitment at 0–400, 950–1450, and 2250–2600 ms, followed by greater emotional-state recruitment at 4400–4850 ms.The authors describe this temporal pattern as complementary evidence of dynamically altered emotion-action coupling during social pain.Source: Time-resolved dynamics reveal an early… · pp. 4–5

Social pain preferentially reveals NSSI-related neurodynamic alterations

• 68.55% accuracy under social pain exceeded 66.45% under physical pain and 61.84% at rest, making social pain the most discriminative condition.The reported performance gradient suggests socially salient distress preferentially exposes NSSI-related differences rather than a generalized alteration across brain states.Source: Social pain preferentially reveals… · p. 4, Table 1 · p. 5

MS5-to-MS3 dynamics differentially link social evaluation to affective outcomes

• In DN− adolescents, MS5→MS3 partially mediated associations between social-evaluation sensitivity and affective outcomes; this pattern was absent in DN+, where MS3→MS5 related to greater negative affect.The affective outcomes included PHQ-9 scores and subjective emotional ratings.Source: MS5-to-MS3 dynamics differentially… · p. 6

Neurodynamic modeling enables individual -level identification of NSSI

• 68.55% accuracy, with an 8.94-percentage-point advantage over the best-performing baseline, was achieved by the proposed neurodynamic model.The model also achieved 69.49% precision, 67.37% recall, 68.08% F1-score, and 68.06% AUC.Source: Neurodynamic modeling enables… · p. 7

Discussion

• NSSI in adolescents with depression is characterized by altered coordination between emotional/interoceptive and action-related brain states, extending region- and network-level accounts toward dynamic interactions.The authors frame the findings as a systems perspective on how rapidly evolving neural states interact over time.Source: Discussion · p. 7

Social pain as a context for revealing NSSI-related vulnerability

• Figure 5 compares DN+ and DN− microstate-occurrence trajectories over 300-ms windows; shaded intervals mark significant group differences.Solid lines denote DN+ and dashed lines DN−; the caption does not state which states differ or when.Source: Figure 5 · p. 8
• Social pain adds rejection, belonging, and self-relevance to systems shared partly with physical pain, increasing demands on emotional appraisal and behavioral regulation.The authors describe social pain as a context that may amplify vulnerability, not as a specific cause of NSSI.Source: Social pain as a context for revealing… · p. 8

Emotion-action decoupling as a neurodynamic mechanism of NSSI

• The proposed dysfunction involves temporal coordination between emotional/interoceptive and action-preparation systems, rather than abnormalities within individual systems alone.Inefficient switching may impair the flexible transformation of socially salient information into adaptive behavioral responses.Source: Emotion-action decoupling as a… · p. 8

Directionality may distinguish emotional integration from behavioral readiness

• The two transition directions suggest distinct roles: action-to-emotion transitions may support affective appraisal, while emotion-to-action transitions may relate to behavioral readiness.The authors caution that neither transition directly represents self-injurious behavior; negative affect may instead become more closely coupled to action-oriented configurations.Source: Directionality may distinguish… · pp. 9–10

Multidimensional analyses support model robustness

• Convergent supplementary checks support framework robustness across optimization, permutation testing, sampling strategies, and hyperparameter settings.Permutation testing found performance above chance (p = 0.03); the authors also report stable convergence and limited sensitivity to hyperparameter selection.Source: Multidimensional analyses support… · p. 10

Sex-related stability of neurodynamic representations

• Performance remained comparable across within-sex evaluations: female Accuracy/F1-score was 67.63%/66.87%, and male Accuracy/F1-score was 65.79%/65.00%.Cross-sex results were less even: Male-to-Female evaluation had Precision of 72.51%, Recall of 53.68%, and F1-score of 61.37%.Source: Sex-related stability of neurodynamic… · p. 10

Independent datasets distinguish generalizability from disease specificity

• On independent MDD datasets, altered bidirectional MS1↔MS3 transitions appeared under eyes-closed HUSM and in MODMA, but not under eyes-open HUSM.The authors interpret these patterns as evidence of disrupted coordination between internally and externally oriented processing modes.Source: Independent datasets distinguish… · pp. 10–11

Participants

• The initial sample comprised 106 adolescents aged 13–18 years; after EEG quality exclusions, 67 DN+ and 34 DN− participants remained.The original DN− group had 39 participants; all participants were diagnosed with MDD, and other psychiatric disorders and high suicide risk were exclusion criteria.Source: Participants · p. 11

Data Acquisition and Preprocessing

• The task used 35 images each depicting physical pain or social rejection, with a five-minute resting recording followed by randomized pain-condition trials.Each trial included a 2-second fixation, 5-second image, 5-second negative-emotion rating on a 1–9 scale, and 1-second black screen.Source: Data Acquisition and Preprocessing · p. 11
• Negative-emotion ratings were collected by keyboard during the assessment segment after each image.Participants rated intensity from 1 (no negative emotion) to 9 (strongest negative emotion).Source: Data Acquisition and Preprocessing · p. 11
• EEG was recorded from 64 electrodes at 500 Hz, and five-second image-stimulation epochs were extracted for analysis.Preprocessing included common-average reference, 1–45 Hz bandpass and 50 Hz notch filters, ICA-based ocular-artifact removal, and bad-channel interpolation.Source: Data Acquisition and Preprocessing · p. 11

Microstate-Based Representation of Global Brain Dynamics

• Global Field Power is computed at each time point as the standard deviation of potentials across electrodes, and maps at GFP peaks are selected for clustering.The peaks represent moments of maximal field stability and highest signal-to-noise ratio.Source: Microstate-Based Representation of… · pp. 11, 13
• Two-stage k-means clustering selects group-level microstate templates from trial-level maps, using GEV and cross-validation criteria to choose the cluster count.The first-stage search tested K = 2–8; the final group templates numbered eight.Source: Microstate-Based Representation of… · p. 13
• Back-fitted templates yield a 2,500-sample microstate sequence for each five-second trial.The group-level templates are fitted to each participant’s preprocessed EEG signals.Source: Microstate-Based Representation of… · p. 13

Multidimensional Data Balancing Strategy

• Stratified 10-fold cross-validation assigns one participant from each gender-by-diagnosis subgroup to testing and the other 36 to training.The balanced sample comprises 10 male and 10 female participants in each diagnostic group.Source: Multidimensional Data Balancing Strategy · pp. 13–14
• Each task block contains 35 trials, with social-pain and physical-pain blocks presented in randomized order alongside a 5-minute resting session.Image order is randomized within each block; participants rate negative emotion after each stimulus.Source: Figure 8 · p. 14
• The balanced model input contains 40 participants, 19 trials per participant, and 2,500 time points per trial.The sampling balances gender, diagnostic class, and trial count.Source: Multidimensional Data Balancing Strategy · p. 14

Blockwise and Randomization Strategies for Disrupting Microstate Temporal Structure

• Sequences are segmented into 50 patches using a window and step size of 50 before temporal perturbation.The perturbations are intended to reduce redundant information and interference from redundant patterns.Source: Blockwise and Randomization Strategies… · p. 14
• Strong permutation shuffles 40–43 patches, effectively destroying the sequence’s specific transition structure.The strategy targets pathological microstate-transition features while reducing redundant information.Source: Blockwise and Randomization Strategies… · p. 14
• Weak permutation shuffles only 1–3 patches, preserving most NSSI-related transition patterns while introducing mild temporal perturbations.Source: Blockwise and Randomization Strategies… · p. 14

Transformer Architecture Based on Patch-Level Feature Modeling

• Patch-level Transformer modeling targets dynamic transitions and captures fine-grained segment dynamics alongside the overall sequence structure.Compared with static occurrence, duration, and coverage features, the approach emphasizes transitions and whole-brain state relationships.Source: Transformer Architecture Based on… · p. 14
• The model embeds patches, prepends a learned CLS patch, and applies a three-layer Patch-Transformer to produce feature vectors.Weighted attention matrices from each layer are retained for later analysis.Source: Transformer Architecture Based on… · p. 15
• Dual positional encodings preserve token-level dependencies before segmentation and patch-level relationships after adding the CLS patch.Source: Transformer Architecture Based on… · p. 15

Multi-loss Joint Optimization

• Disease-specific loss is designed to distinguish DN+ from DN− participants by identifying features unique to each condition.The paper presents it as supporting the model’s discriminative capability.Source: Multi-loss Joint Optimization · p. 15
• Domain-adversarial loss targets domain-invariant features to improve generalization across domains and reduce individual-difference effects.It uses source- and target-domain samples, a feature extractor, a domain discriminator, and a gradient-reversal layer.Source: Multi-loss Joint Optimization · p. 15
• Consistency learning distinguishes context vectors from original sequences and strongly permuted sequences using pseudo-labels and binary cross-entropy.A single-layer unidirectional LSTM produces the context vectors.Source: Multi-loss Joint Optimization · p. 15
• Contrastive learning uses the original sequence as anchor, weakly permuted sequence as positive, and strongly permuted sequence as negative to sharpen group discrimination.The approach employs InfoNCE loss on context vectors.Source: Multi-loss Joint Optimization · p. 15
• The total objective directly sums disease-specific, domain-adversarial, consistency, and contrastive losses to jointly optimize the model.The corresponding weighting factors are α, β, δ, and θ.Source: Multi-loss Joint Optimization · p. 16

• Cosine similarity between learned neural representations and stimulus image features was compared across groups and pain conditions using two-way mixed ANOVA.Source: Statistical Analysis · p. 16
• Paired t-tests on cross-validation-fold attention matrices evaluated microstate transition-weight differences between disease and sex subgroups.Source: Statistical Analysis · p. 16
• Behavioral measures were examined with correlation and mediation analyses, while temporal-validity microstate features were compared between groups using independent two-sample t-tests.Source: Statistical Analysis · p. 16
• Figures mark significance at p ≤ 0.05, p ≤ 0.01, and p ≤ 0.001 with one, two, and three asterisks, respectively.Source: Statistical Analysis · p. 16