Multi-Frequency EEG Connectomics Uncovers Insula-Network Subtypes in Somatic Symptom Disorder

Multi-Frequency EEG Connectomics Uncovers Insula-Network Subtypes in Somatic Symptom Disorder | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Multi-Frequency EEG Connectomics Uncovers Insula-Network Subtypes in Somatic Symptom Disorder Nan Yan, Shuzhi Zhao, Chongyuan Lian, Xue Shi, Ge Dang, Zi’an Pei, and 7 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7971302/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Substantial clinical heterogeneity in Somatic Symptom Disorder (SSD) limits treatment efficacy. Here， this study propose a data-driven framework using neurophysiology information to identify distinct patient subtypes. A contrastive variational autoencoder with Gaussian mixture modeling (CVAE-GM) was developed using resting-state EEG connectomics from a discovery cohort of 1,419 patients with SSD. The derived subtypes were clinically correlated with symptom dimensions and validated for reproducibility in an independent external cohort (n=530). We identified three robust and reproducible subtypes, characterized by dominant connectivity in somatomotor, central executive, or limbic networks with insula area. Each neurophysiological subtype was significantly associated with distinct clinical profiles (somatic symptoms, adverse cognitive, and negative emotions). The subtyping model demonstrated high reproducibility in the validation cohort (accuracy = 0.85; AUC = 0.87). This EEG-based framework provides a validated, neurobiological basis for stratifying SSD patients. It transcends symptom-based nosology, enabling a precision medicine approach for developing targeted interventions. Health sciences/Biomarkers/Diagnostic markers Biological sciences/Computational biology and bioinformatics/Machine learning Health sciences/Diseases/Psychiatric disorders Biological sciences/Biological techniques/Electrophysiology/Electroencephalography – EEG Biological sciences/Neuroscience/Somatosensory system/Insula Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Somatic Symptom Disorder (SSD) imposes a substantial burden on healthcare systems, driven by a cycle of distressing physical symptoms and the excessive cognitive and emotional responses they provoke 1, 2 . Its high prevalence, affecting up to a third of patients in general hospitals 3 , underscores its substantial impact on public health. Although validated psychometric scales such as the Patient Health Questionnaire-15 (PHQ-15) 4 , the Somatic Symptom Disorder-B Criteria Scale (SSD-12) 5 , and the Neuro-11 6 have been developed, the current categorical framework aggregates clinically heterogeneous presentations, which likely reflect distinct and overlapping neurobiological mechanisms 7, 8 . This heterogeneity results in more than 40% of patients failing to respond adequately to standard pharmacological and psychotherapeutic interventions 9 . These limitations underscore the need for a precision-medicine approach that integrates computational modeling with neurophysiological insights to enable personalized patient stratification and clinically interpretable subtyping. SSD may be conceptualized as a disorder of hierarchical predictive coding, wherein aberrant precision weighting of interoceptive prediction errors amplifies innocuous bodily sensations into persistent and distressing symptoms 10 . Specifically, excessive precision of somatic priors prevents the appropriate updating of internal models, resulting in sustained prediction error signals that manifest as clinically unexplained symptoms 11, 12 . These computational disruptions are thought to engage mechanisms of central sensitization as a transdiagnostic process across multiple somatic domains. Central sensitization is hypothesized to be mediated by maladaptive, context-dependent functional alterations in brain regions that mediate interoceptive 10 , affective, and sensory processing, including the somatosensory cortex, prefrontal cortices, insula cortex, anterior cingulate cortex, thalamus, and brainstem. In this context, neuroimaging-based subtyping represents a promising avenue for identifying neurobiologically homogeneous patient subgroups, paving the way for the development of pathophysiological biomarkers and therapeutic interventions. Electroencephalography (EEG) offers a powerful means to characterize the tempo-spatial dynamics of neural processes implicated in the clinical heterogeneity of SSD. Previous studies employing spectral power analysis 13, 14 , phase synchronization 14 , and functional connectivity 15 , along with machine learning architectures, have differentiated SSD patients from healthy controls. While previous supervised machine learning models have achieved high accuracy in distinguishing SSD patients from healthy controls, they fail to address the critical issue of patient-level heterogeneity, a key factor limiting treatment efficacy. This underscores a significant gap: the need for methods that can stratify patients rather than merely classify them. Self-supervised learning offers a promising solution by uncovering latent, data-driven structures without reliance on predefined diagnostic labels 16, 17 . Therefore, the objective of this study was to develop and validate a computational framework using a contrastive variational autoencoder (CVAE) applied to large-scale EEG connectomics. However, applying these subtyping methods faces significant challenges 18 . Current models often analyze single-frequency bands, ignoring crucial cross-frequency dynamics implicated in SSD, a critical aspect of the neurophysiological principle underlying complex cognitive processes. They also focus on group-level contrasts, failing to capture critical within-group heterogeneity. Finally, their inherent lack of interpretability acts as a major barrier, preventing the translation of model features into mechanistic insights about symptom drivers. Although post-hoc attribution methods (e.g., SHAP, Integrated Gradients) can identify salient features, they fall short of elucidating the causal or theoretical relationships between neural dynamics and symptom dimensions. To address these limitations, we introduce a computational framework that identifies neurophysiological subtypes of SSD by integrating multi-frequency EEG connectivity with self-supervised learning. Our model leverages multi-band cross-attention to capture hierarchical neural dynamics, a key aspect of neural multiplexing. Contrastive learning then identifies patient subgroups as deviations from a normative neural manifold, while a Gaussian mixture model accommodates the disorder's profound heterogeneity. The framework successfully identifies robust, reproducible subtypes whose distinct connectivity patterns are significantly associated with specific symptom dimensions. This integrative approach represents a paradigm shift from phenomenological classification toward a mechanistically-informed psychiatric nosology, offering a path to enhanced diagnostic precision and the design of individualized interventions. [ Insert Figure 1 near here ] Results Can the integration of cross-frequency EEG connectome analysis with contrastive learning methods identify neurobiologically distinct subtypes within SSD? We investigated whether our conditional variational autoencoder with Gaussian mixture modeling (CVAE-GM) could identify biologically meaningful subtypes within the large SSD electroencephalography dataset (n = 1,419). To determine the optimal cluster number, we applied the Calinski-Harabasz index 19 alongside distributional analysis of within-cluster scalp network strength. While this index suggested either two or three clusters (Supplementary Fig. 1A), distributional analysis 20 revealed that the two-cluster solution produced multimodal distributions within clusters, indicating residual heterogeneity (Supplementary Fig. 1B). Conversely, the three-cluster solution yielded unimodal, well-concentrated distributions (Supplementary Fig. 1C), leading us to select k = 3 for subsequent analyses. [ Insert Figure 2 near here ] The CVAE-GM model identified three distinct subtypes with minimal overlap (mean = 1.5%), achieving a 62.5% improvement in cluster separability over standard variational autoencoders (4.1% overlap). Features from theta and beta frequency bands provided the highest discriminatory power, highlighting the importance of frequency-specific neural dynamics in subtype differentiation. The three subtypes were characterized by distinct network dominance patterns: sensorimotor network (SMN)-dominant, central executive network (CEN)-dominant, and limbic network (LN)-dominant (Fig. 2), each of which is defined by the distinct functional activation pattern represented by its cluster centroid. To assess intra-subtype heterogeneity, we examined continuous variation in connectivity patterns as a function of distance from each cluster's centroid. This analysis is predicated on the principle that a continuous representation of an individual's similarity to a subtype centroid provides a finer-grained characterization of inter-individual differences than categorical assignment alone (See Supplementary Appendix). Individual connectivity maps showed progressive divergence from centroid prototypes with increasing distance across all three subtypes (Fig. 3A–C). Kullback-Leibler and Jensen-Shannon divergence analyses confirmed substantial between-subtype differentiation with low cross-modal concordance (Fig. 3D). [ Insert Figure 3 near here ] Internal validation confirmed subtype robustness across multiple approaches, as proposed by 18 . SigClust analysis 21 established statistical significance for all three subtypes (p_FDR < 0.001). Consistency metrics demonstrated excellent clustering performance (V-measure: 0.87; adjusted Rand index: 0.88; normalized mutual information: 0.86; silhouette coefficient: 0.84) (Supplementary Fig. 1c). Overall, these analyses suggest that the identified neuro-subtypes are internally valid and likely reflect the heterogeneity of the EEG data. Note that because of the limitations of the testing methods, these analyses provide evidence but do not guarantee the separability of the identified subtypes. In the absence of ground truth, internal validation of clustering results remains challenging, thus necessitating external validation using metrics that were not employed in the clustering analysis. Do neurobiologically distinct subtypes exhibit specific clinical profiles? The Canonical Correlation Analysis (CCA) 22 was employed to assess the multivariate relationship between subtype-specific functional connectivity density (FCD) patterns and the clinical dimensions of the Neuro-11 questionnaire. In this analysis, we found significant specific mappings (See Figure 4): The SMN-dominant subtype was characterized by elevated FCD within the SMN, which was significantly and positively correlated with somatic symptom scores on the Neuro-11 scale (r = 0.42, p < 0.001). Similarly, the CEN-domain subtype exhibited increased FCD within the CEN, which was negatively correlated with scores on the negative life events dimension (r =–0.38, p < 0.001). Finally, the LN-domain subtype exhibited heightened FCD within the LN, which was positively correlated with emotion dimension scores (r = 0.36, p = 0.002). In contrast, no other networks showed significant clinical associations (all p > 0.05). [ Insert Figure 4 near here ] Shared and distinct neural substrates across SSD subtypes To validate model robustness, we incorporated age- and sex-matched healthy controls (n = 112). The CVAE-GM model maintained effectiveness, identifying four clusters with 8.5% mean overlap. Healthy controls formed a distinct cluster characterized by default mode network dominance in theta and beta bands, consistent with established neurophysiology. The three patient subtypes remained stable (Supplementary Fig. 2). FDR-Group-level comparisons between each subtype and controls revealed distinct neural substrates. The SMN-dominant subtype exhibited hyperactivation in the anterior cingulate cortex (ACC), ventrolateral prefrontal cortex (vlPFC), insula, and premotor cortex, suggesting overactive interoceptive-sensorimotor integration. The CEN-dominant subtype showed aberrant activity in the dorsolateral prefrontal cortex (DLPFC), precuneus, and insula, reflecting dysregulated inhibitory control. The LN-dominant subtype displayed hyperactivity in the thalamus, posterior cingulate cortex (PCC), and occipitotemporal junction, indicating limbic-autonomic dysfunction. [ Insert Figure 5 near here ] Notably, insula dysfunction was observed across all subtypes, suggesting a core pathophysiological hub. Connectivity analyses revealed subtype-specific insula coupling patterns: altered insula–anterior cingulate connectivity (SMN-dominant), disrupted insula–dorsolateral prefrontal connectivity (CEN-dominant), and abnormal insula–thalamus connectivity (LN-dominant) (Fig. 5B). These aberrant connectivity pathways reinforce the notion that dysfunctional insula interactions with specific brain regions may represent a central neurobiological mechanism underlying different symptomatic manifestations of SSD. Independent validation confirms subtype reproducibility Our analyses revealed a critical dissociation: while distinct data modalities independently identified robust subtypes with convergent clinical profiles. A central question, therefore, is whether these findings are specific to the cohort under investigation or reflect broader patterns observable across diverse populations. We validated our findings using an independent SSD dataset (n = 530). Participants were assigned subtype labels based on predominant clinical dimensions (somatic, negative emotion, or adverse events) from Neuro-11 assessments. A comparative analysis of state-of-the-art self-supervised learning methods demonstrated that all approaches achieved high classification accuracy, confirming the underlying heterogeneity of SSDs. Our CVAE-GM framework achieved superior performance (accuracy: 0.85; AUC: 0.87; sensitivity: 0.86; 1-specificity: 0.88) (Table 1), validating both the existence of distinct neurophysiological subtypes and the effectiveness of our methodology in capturing this heterogeneity. [ Insert Table 1 near here ] Discussion In this study, we propose a framework for stratifying individuals with SSD based on latent neural signatures associated with symptom heterogeneity. By integrating contrastive variational autoencoders, Gaussian mixture modeling, and multi-band EEG functional connectivity, our model identified three distinct SSD subtypes characterized by distinct patterns of neural dysregulation that corresponded to specific symptom dimensions measured by the Neuro-11 scale. These results move beyond traditional diagnostic categories by establishing a mechanistic basis for data-driven subtyping, offering a concrete step toward precision psychiatry in functional somatic syndromes. Neurobiological mechanisms underlying SSD heterogeneity. Our findings indicate that symptom heterogeneity in SSD is underpinned by distinct patterns of CFC dysfunction across large-scale networks. In particular, the SMN-dominant subtype exhibited hyperconnectivity within sensorimotor circuits (P<0.001), characterized by θ-β CFC enhancement between the insula and key regions involved in sensorimotor integration, including anterior cingulate cortex (ACC, modulation index: 0.65±0.05 vs 0.45±0.07), premotor cortex (PMC, 0.60±0.06 vs 0.44±0.07), and precuneus (0.60±0.05 vs 0.43±0.06; all P < 0.001). This CFC enhancement was significantly correlated with somatic symptom severity (r = 0.46, P < 0.001). This pattern is consistent with disrupted GABA-mediated inhibitory control within sensorimotor circuits 23 , compromising the hierarchical predictive coding processing required for accurate body-state estimation. Aberrant θ-β coupling disrupts this balance, leading to misaligned prediction-error signaling whereby innocuous interoceptive stimuli are misinterpreted as salient bodily threats 24 . This is consistent with models of dysfunctional Bayesian inference, wherein overly precise prediction errors amplify normal bodily signals into exaggerated somatic experiences 25 . Functionally, the observed insula-ACC-PMC-precuneus hyperconnectivity implicates a dysfunctional network responsible for integrating interoceptive awareness (insula), emotional salience (ACC), motor intentions (PMC), and self-referential bodily representation (precuneus). Such dysregulation may contribute to a fragmented body schema and an exaggerated sense of symptom authenticity. Importantly, this neurophysiological signature helps explain why patients in the SMN-dominant subtype experience their symptoms as subjectively real despite the absence of identifiable peripheral pathology. Rather than reflecting mere emotional amplification or executive misattribution, the findings point to a fundamental miscalibration of internal sensorimotor models, wherein aberrant oscillatory dynamics generate persistent and biologically encoded prediction errors 26, 27 . The CEN-dominant subtype exhibited increased connectivity density within fronto-parietal circuits (P < 0.001) and enhanced dlPFC-anterior insula coupling (0.63±0.05 vs 0.46±0.06, P < 0.001). These network alterations were inversely correlated with the negative event reactivity of Neuro-11 (r=-0.38, P < 0.001). This “hyperregulation” state represents a pathological configuration in which excessive top-down cognitive control disrupts adaptive interoceptive processing 10 , producing a form of cognitive override that amplifies, rather than attenuates, somatic symptoms through hypervigilant monitoring and catastrophic misinterpretation of bodily sensations 28, 29, 30 . The observed neural hyperactivation is therefore paradoxical: it represents an energetically expensive, dysfunctional process where cognitive engagement with interoceptive signals perpetuates symptomatology instead of providing regulatory control 31 . The LN-dominant subtype exhibited significantly increased functional connectivity within limbic circuits compared to healthy controls ( P < 0.001), with particularly robust enhancement along the thalamo-insula pathway (connectivity strength: 0.61±0.06 vs 0.44±0.06, P < 0.001). This hyperconnectivity was positively correlated with emotional symptom severity of the Neuro-11 scale (r = 0.36, p = 0.002), indicating a convergent limbic-somatic pathophysiology 30, 32, 33 . Thalamo-insula hyperconnectivity appears to drive a limbic-mediated pathway of symptom amplification 34 . Within this circuit, excessive thalamic input to the posterior insula likely imbues normal bodily sensations with aberrant threat salience 35 . This mechanism stands in stark contrast to the top-down cognitive hypercontrolled of the CEN-dominant subtype, explaining a phenotype in which patients recognize the irrationality of their symptoms yet cannot override the powerful, affectively charged interoceptive experience 36 . The insula functions as a pathological convergence hub driving subtype-specific symptom generation in somatic symptom disorders. Across all three SSD subtypes, we observed consistent abnormalities in insular connectivity, suggesting this region as a central hub in the pathophysiology of SSD. This convergence aligns with the insula’s unique anatomical and functional properties: it receives direct ascending interoceptive input via lamina I spinothalamic pathways and contains Von Economo neurons that support rapid interoceptive–cognitive integration 24, 37 . These features position the insula as a key cortical interface through which bodily sensations acquire conscious representation and affective significance 10 . Notably, the clinical heterogeneity of SSD appears to arise from subregion-specific patterns of insular dysconnectivity. In SMN-dominant patients, posterior insular hyperconnectivity with somatosensory cortices is associated with heightened bodily awareness and amplified sensorimotor salience. In LN-dominant cases, enhanced anterior insula–amygdala coupling correlates with emotional symptom severity, implicating dysregulated limbic integration of interoceptive signals. In CEN-dominant subtypes, increased connectivity between the dorsal anterior insula and dlPFC reflects maladaptive cognitive monitoring of internal bodily states 33, 38 . These findings support a tripartite model of insular dysfunction that reconciles SSD heterogeneity within a unified neurobiological framework: hyperactive posterior insula processing amplifies somatic sensations; dysregulated mid-insula integration perturbs salience attribution; and aberrant anterior insula coupling with executive or limbic circuits determines whether symptoms manifest as cognitive hypervigilance, emotional dysregulation, or sensorimotor distortion 39 . The insula thus emerges as both the proximal generator of symptom experience and a promising therapeutic target, where subtype-specific modulation of insular circuitry may offer a promising pathway for restoring adaptive interoceptive processing 10 and reducing symptom burden across heterogeneous SSD profiles. Neurobiologically-informed EEG subtyping transcends symptom-based taxonomies to enable precision psychiatry in somatic symptom disorders. Current psychiatric classifications yield treatment failure rates of 30-40% in SSD, reflecting their reliance on symptom-based categorization that ignores underlying pathophysiology 40 . While standardized clinical instruments such as the PHQ-15 and SCL-90 achieve moderate diagnostic accuracy, their dependence on subjective self-report limits their capacity to capture the multidimensional neurophysiological heterogeneity that drives symptom expression 3, 8, 41 . The Hierarchical Taxonomy of Psychopathology (HiTOP) represents a significant conceptual advance by modeling psychopathology along dimensional spectra, and tools such as the PID-5 have achieved improved predictive validity (AUC = 0.81-0.84) across both categorical and dimensional assessments 42 . However, HiTOP remains fundamentally symptom-centric, offering limited insight into the neural mechanisms that are essential for biologically informed, mechanism-targeted interventions. Our proposed EEG-based machine learning framework addresses these limitations by quantifying the neural substrates underlying SSD heterogeneity. The CVAE-GM architecture with multi-frequency contrastive learning achieved superior classification performance (balanced accuracy = 0.85, AUC = 0.87) compared to symptom-based approaches, while identifying three distinct neurofunctional subtypes with reproducible connectivity signatures. Unlike conventional diagnostic systems that describe phenomenology, therefore, our framework enables the direct identification of mechanistic subtypes to enhance diagnostic precision and inform individualized therapeutic strategies for SSD. Methodological advances and comparative performance. Our CVAE-GM framework exhibited substantial methodological advances over conventional approaches, achieving a 61.1% improvement in subtype stability (cluster separability silhouette coefficient: 0.87 vs 0.54 for Gaussian mixture models, P < 0.001) and enhanced discriminative power (AUC: 0.87, 95% CI: 0.85-0.93 vs 0.77, 95% CI: 0.71-0.81 for VAE). A central advantage of our framework lies in its contrastive learning architecture, which captures the distributed and frequency-specific nature of neural representations via multiband latent embeddings. This enables the identification of population-level oscillatory patterns that elude traditional univariate or frequency-agnostic models. Performance gains were further driven by a novel coupling consistency loss function that imposes neurobiologically informed constraints on cross-frequency interactions, mirroring neural ensemble synchronization mechanisms observed in healthy cortical circuits. To our knowledge, this represents the first participant-level clustering method that utilizes multiband EEG features in conjunction with Gaussian mixture contrastive learning to derive biologically informed subtypes. Importantly, it addresses a fundamental limitation of existing models that treat EEG as static connectivity matrices, instead modeling the dynamic and oscillatory nature of neural communication that underlies complex psychiatric phenotypes. The resulting subtypes are not only statistically robust but also neurobiologically plausible, exhibiting connectivity signatures consistent with established principles of cortical organization and cross-frequency coupling. Limitations and future directions. Several limitations of this study should be acknowledged. First, while our resting-state EEG approach captures static functional connectivity patterns, it may not fully reflect dynamic or context-dependent neural dysfunctions. Future studies should incorporate task-based paradigms targeting interoceptive attention, emotional regulation, and executive control, as well as perturbational techniques such as TMS–EEG or neurofeedback, to establish causal relationships between network dynamics and behavioral phenotypes 43 . Second, the cross-sectional design and demographically homogenous sample (n = 150, exclusively Chinese) limit generalizability. Longitudinal studies tracking connectivity trajectories over time, along with multi-site replications in ethnically diverse populations, are essential to evaluate cultural influences and ensure cross-populational robustness. Third, combining EEG subtypes with high-resolution fMRI, polygenic risk scores, and peripheral biomarkers could yield comprehensive neurobiological profiles. And emerging computational advances, including physics-informed neural networks incorporating biophysical constraints and federated learning enabling multi-site validation, could offer promising avenues for developing diagnostic and therapeutic algorithms that transform SSD from heterogeneous diagnostic categories into distinct, biologically-defined conditions 44, 45 . Conclusions and translational implications. We present a novel framework that elucidates the neural underpinnings of symptom dimensions in SSD using resting-state EEG, identifying three robust, neurobiologically-distinct subtypes. The clinical validity of these subtypes is supported by their strong correspondence with Neuro-11 symptom dimensions and distinct spectral patterns, particularly in cross-frequency coupling dynamics. These findings establish a data-driven, mechanistically grounded approach for stratifying SSD, providing significant insights into translating neurofunctional subtypes into personalized diagnostic and therapeutic strategies. Methods Participants A total of 1,949 individuals diagnosed with somatic symptom disorder (SSD) and 112 healthy controls were recruited across multiple clinical departments at Shenzhen People's Hospital. The SSD cohort was stratified into an internal training dataset (n = 1,419; mean age 48.11 ± 15.70 years) recruited from Neurology departments, and an external validation dataset (n = 530; mean age 52.23 ± 16.11 years) recruited from Cardiology, Gastroenterology, Endocrinology, and Traditional Chinese Medicine departments. Healthy controls (n = 112; mean age 42.03 ± 13.72 years) were age-matched to the patient cohorts. To enhance ecological validity, both inpatient and outpatient populations were included. All participants provided informed consent before their participation, and procedures were approved by the institutions’ Institutional Review Boards. All patients met the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) 46 criteria for SSD, characterized by persistent somatic symptoms accompanied by excessive health-related thoughts, feelings, and behaviors. Diagnostic assessments were conducted by experienced psychiatrists using a battery of standardized clinical instruments, including the Hamilton Depression Rating Scale (HAMD) 47 , the Hamilton Anxiety Rating Scale (HAMA) 48 , the Pittsburgh Sleep Quality Index (PSQI) 49 , and the Neuro-11 questionnaire 6 . Crucially, the final diagnosis was based on comprehensive clinical judgment rather than predetermined scale cutoff scores. Exclusion criteria included: a history of severe psychiatric disorders (e.g., schizophrenia, bipolar disorder, major depressive disorder); current use of psychotropic medications or neuromodulation treatments (e.g., repetitive transcranial magnetic stimulation); significant cognitive impairment (e.g., dementia), language barriers, or impaired consciousness; substance abuse, suicidal risk, or recent cranial trauma or neurological disorders (e.g., epilepsy, Parkinson’s disease, amyotrophic lateral sclerosis); pregnancy or lactation; and any other medical or ethical considerations deemed unsuitable for study participation. EEG Acquisition and Signal Processing Resting-state EEG data (8 min eyes-open, 8 min eyes-closed) were acquired from 64 channels in a 10-10 configuration using a BrainAmp DC amplifier (Brain Products GmbH) at a sampling rate of 1,000 Hz. Impedances were kept below 5 kΩ, with FCz serving as the reference and AFz as the ground. The raw data were processed using a pipeline incorporating EEGLAB 50 and MNE-Python 51 . The pipeline included 50 Hz notch and 1-45 Hz band-pass filtering, manual rejection of non-stereotyped artifacts, bad channel interpolation, and ICA-based removal of ocular and cardiac artifacts. The cleaned data were subsequently re-referenced to the common average. From these signals, we extracted the power spectral density (PSD) via FFT for six frequency bands: delta (1-4 Hz), slow theta (4-6 Hz), fast theta (6-8 Hz), alpha (8-13 Hz), beta (13-30 Hz), and gamma (30-45 Hz). Finally, cortical source activity was estimated using a standard forward and inverse modeling approach. Identification of SSD Subtypes via a CVAE-GM Model We propose a novel deep learning model, the CVAE-GM, to identify distinct subtypes of SSD from neurophysiological data. This model combines a Conditional Variational Autoencoder (CVAE) with a Gaussian Mixture (GM) model in its latent space. This architecture addresses the limitations of standard variational autoencoders, which assume unimodal latent distributions and inadequately capture heterogeneous clinical populations. The model integrates three key components: (1) A multi-band contrastive learning module that enforces feature invariance across frequency ranges while maintaining individual-specific pathophysiological signatures; (2) A Gaussian mixture latent space that replaces the standard unimodal Gaussian prior with a multimodal distribution, enabling direct modeling of distinct neural phenotypes; (3) An integrated training framework that embeds contrastive learning within the CVAE-GM optimization loop, yielding a latent space optimized for generation, clustering, and cross-frequency dynamics capture. (see Supplementary Materials for architectural and training details). 1) Multi-band Contrastive Learning Module . To generate robust embedding feature representations, our model employs a contrastive learning objective across frequency bands. This module learns an embedding space where features from low-frequency (delta, theta) and high-frequency (beta, gamma) bands of the same individual are pulled closer, while those from different individuals are pushed apart. This process enforces the learning of features that are invariant across frequency ranges but specific to an individual's pathophysiology, thereby enhancing subtype separation. 2) Gaussian Mixture Latent Space. Standard VAEs typically assume a unimodal Gaussian prior 52, 53 , an oversimplification for heterogeneous clinical populations like SSD. To address this, our model's latent space is regularized by a Gaussian Mixture Model prior. This allows the encoder to map individuals to distinct clusters within the latent space, directly modeling the multimodal distribution of neural phenotypes and improving the accuracy of subtype assignment. 3) Integrated Architecture for Subtyping. The final architecture embeds the contrastive learning objective within the CVAE-GM training loop. This unified approach ensures that the model learns a latent space that is not only well-structured for generation (via the VAE) and clustering (via the GMM) but also enriched with features that capture cross-frequency dynamics (via contrastive learning). The result is a robust framework for data-driven subtyping that yields interpretable neural clusters to inform precision medicine. Subtype validation framework using the comprehensive metrics Subtype Stability Analysis . A stable clustering solution should be robust to small perturbations of the dataset, such as removing a subset of subjects from the clustering procedure. We verified the stability of the identified subtypes by running the CVAE-GM model 100 times with different random initializations. The consistency of the resulting subject assignments was measured by calculating the average cluster overlap, which is a quantitative metric used to assess the stability and robustness of a clustering solution. We used the Overlap Coefficient (OVL) to quantify the overlap between subtype connectivity distributions and the standard normal distribution, thereby detecting residual heterogeneity or multimodality within subtypes. Estimation followed the parametric binormal and nonparametric kernel approach described by Franco‐Pereira et al. 54 . Subtype validity and robustness Analysis. To comprehensively evaluate the validity and robustness of the identified SSD subtypes, we employed a suite of both external and internal validation metrics. Individual-centroid divergence was measured using Kullback-Leibler (KL) and Jensen-Shannon divergences (JSD), capturing directional bias and overall similarity, respectively. For each subject, KL and JSD were calculated against the three centroids, and group-level differences were assessed using nonparametric rank tests with FDR correction. External validation was performed by comparing our model's subtype assignments to core dimensions of Neuro-11. We quantified the concordance using three distinct metrics: the Normalized Mutual Information (NMI), which measures the mutual dependence between partitions; the Adjusted Rand Index (ARI), which calculates similarity while correcting for chance; and the V-measure, which balances cluster homogeneity and completeness. For internal validation, which assesses the intrinsic structure of the data without external labels, we calculated the average silhouette coefficient. This metric quantifies the degree of cluster definition by measuring the ratio of inter-cluster separation to intra-cluster cohesion. A high degree of convergence across these diverse metrics would provide strong evidence for a meaningful and reliable subtyping solution. Significance of the subtypes. To formally assess the statistical significance of the identified subtype structure, we employed the Significance of Clustering (SigClust) method 55 . This method formally evaluates the null hypothesis that the observed data originates from a single multivariate Gaussian distribution, as opposed to a Gaussian mixture. Through Monte Carlo simulation, SigClust calculates a p-value for this null hypothesis. A significant p-value indicates that the single-Gaussian model is a poor fit, thus providing statistical support for the presence of a genuine cluster structure in the data. Additionally, we constructed a continuous similarity gradient within each subtype. Specifically, each individual’s functional connectivity pattern was compared with the subtype centroid (using correlations of connectivity density maps), and subjects were ordered by similarity to form a gradient from “center” to “periphery.” This approach, adapted from prior work on continuous subtyping 15, 56 , provides a finer characterization of intra-subtype heterogeneity. Detecting the heterogeneous patterns of functional brain activity among subtypes To elucidate the neurofunctional significance of the SSD subtypes identified by the CVAE-GM model, we performed a series of functional activity and Graph-based analyses on the source-reconstructed EEG data. First, the minimum norm estimate (MNE) method was applied to reconstruct resting-state EEG signals in source space, mapping electrophysiological activity to the cortical surface to obtain region-level time-series activity patterns. We applied the seven-network Yeo parcellation 57 to investigate alignment between CVAE-GM patterns and established functional architectures. Neural dynamics were characterized through phase-space reconstruction and attractor analysis 58 to detect transitional activation patterns indicative of unstable neural states. Graph-theoretical analysis computed intra-network connection density for each functional network, defined as the ratio of existing to possible connections within networks. This quantified network integration and local specialization across subtypes. Group comparisons with healthy controls employed two-sample t-tests with false discovery rate correction (p < 0.05) to identify subtype-specific neural alterations. Association Analysis between Subtypes and Clinical Assessments We used Canonical Correlation Analysis (CCA) 22 to map the CVAE-GM-derived neural subtypes onto clinical symptom dimensions. The analysis tested for significant multivariate correlations between the neural latent features and scores from three domains of the Neuro-11 questionnaire 6 : somatic symptoms, negative affect, and adverse life events. By interpreting the significant canonical loadings, we identified the specific clinical profiles most characteristic of each neural subtype, thereby providing insight into the neural basis of clinical heterogeneity in SSD. External Validation on an Independent Dataset. To assess the generalizability and robustness of the proposed CVAE-GM model, we conducted a comprehensive evaluation on a large, independent EEG dataset of individuals with SSD. The model was trained exclusively on the source dataset and tested on the independent cohort. Ground-truth labels for classification were assigned based on the dominant symptom dimension identified by the Neuro-11 questionnaire. We established a unified comparative framework to benchmark CVAE-GM against nine representative self-supervised learning models, including VAE 59 , MPVAE 37 , POE-VAE 60 , and their GM-integrated counterparts, as well as three state-of-the-art models (NSSINet 41 , MMM 61 , LaBraM 62 ). To ensure a fair and reproducible comparison, all models were trained using a five-fold cross-validation scheme with identical input features (multi-band functional connectivity), optimizer, learning rate, batch size, and a maximum of epochs with an early stopping policy. All experiments were executed within a standardized hardware and software environment. Model performance was quantified using Accuracy (ACC), Area Under the Curve (AUC), Sensitivity, and Specificity. Declarations Code availability All custom source code developed for this study, including the implementation of the conditional variational autoencoder with Gaussian mixture prior (CVAE-GM), is publicly available at https://github.com/zhaoshuhzi/CVAE-GM under the MIT License. The repository includes scripts, model definitions, and documentation necessary to reproduce the analyses and results presented in this work. Acknowledgements We thank all members of the Yan and Guo laboratories for their valuable discussions and technical assistance throughout this project. We are also grateful to Prof. Dezhong Yao and Prof. Lan Wang for their insightful comments on the analysis and interpretation of the data. This work was supported by the National Natural Science Foundation of China (U23B2018, 62271477, 82371471), the Shenzhen Science and Technology Program (JCYJ20220818101217037, KQTD20200820113106007, JCYJ20240813151223031), and the Shenzhen Natural Science Foundation. Author contributions Nan Yan and Shuzhi Zhao conceived the study. Nan Yan, Shuzhi Zhao, and Yi Guo designed the methodology and analysis plan. Xin Jiang, Chongyuan Lian, and Xue Shi conducted data collection and preprocessing. Nan Yan and Shuzhi Zhao developed the machine learning models and performed statistical analysis. Nan Yan, Hanjun Liu, and Shuzhi Zhao interpreted the results. Nan Yan drafted the initial manuscript. All authors critically reviewed and approved the final version. Nan Yan and Yi Guo had full access to all study data and verified its integrity. All authors had final responsibility for the decision to submit for publication. Competing interests The authors declare no competing interests. References Löwe B , et al. Persistent physical symptoms: definition, genesis, and management. The Lancet 403 , 2649-2662 (2024). Löwe B , et al. Somatic symptom disorder: a scoping review on the empirical evidence of a new diagnosis. Psychological Medicine 52 , 632-648 (2022). Tozzi L , et al. Personalized brain circuit scores identify clinically distinct biotypes in depression and anxiety. Nature medicine 30 , 2076-2087 (2024). Kroenke K, Spitzer RL, Williams JB. The PHQ-15: validity of a new measure for evaluating the severity of somatic symptoms. Psychosomatic medicine 64 , 258-266 (2002). Toussaint A , et al. Development and Validation of the Somatic Symptom Disorder-B Criteria Scale (SSD-12). Psychosom Med 78 , 5-12 (2016). Zeng S , et al. Neuro-11: a new questionnaire for the assessment of somatic symptom disorder in general hospitals. General Psychiatry 36 , (2023). Bogaerts K , et al. Brain mediators of negative affect-induced physical symptom reporting in patients with functional somatic syndromes. Translational psychiatry 13 , 285 (2023). Kas MJ , et al. Precision psychiatry roadmap: towards a biology-informed framework for mental disorders. Molecular psychiatry , 1-10 (2025). Hallberg H, Maroti D, Lumley MA, Johansson R. Internet-delivered emotional awareness and expression therapy for somatic symptom disorder: one year follow-up. Frontiers in Psychiatry 15 , 1505318 (2025). Zhang J , et al. Cortical and subcortical mapping of the human allostatic-interoceptive system using 7 Tesla fMRI. Nat Neurosci , (2025). Seymour B, Crook RJ, Chen ZS. Post-injury pain and behaviour: a control theory perspective. Nature reviews Neuroscience 24 , 378-392 (2023). Zhang Y, Chen ZS. Harnessing electroencephalography connectomes for cognitive and clinical neuroscience. Nat Biomed Eng 9 , 1186-1201 (2025). Grabowski L B, Chelstowski M, Hiszpanska M, Laszewska K, Lewandowska M, Milner R. Abnormalities in the absolute power of Delta and Alpha rhythms in the frontal lobe of patients suffering from psychosomatic disorders. Psychiatria polska , 1-12 (2024). Hong JK, Park HY, Yoon IY, Jang YE. Longitudinal qEEG changes correlate with clinical outcomes in patients with somatic symptom disorder. J Psychosom Res 151 , 110637 (2021). Novelli L, Razi A. A mathematical perspective on edge-centric brain functional connectivity. Nat Commun 13 , 2693 (2022). Liu B , et al. Self-supervised learning reveals clinically relevant histomorphological patterns for therapeutic strategies in colon cancer. Nat Commun 16 , 2328 (2025). Rafiei MH, Gauthier LV, Adeli H, Takabi D. Self-Supervised Learning for Electroencephalography. IEEE Trans Neural Netw Learn Syst 35 , 1457-1471 (2024). Brucar LR, Feczko E, Fair DA, Zilverstand A. Current approaches in computational psychiatry for the data-driven identification of brain-based subtypes. Biological psychiatry 93 , 704-716 (2023). Wang X, Xu Y. An improved index for clustering validation based on Silhouette index and Calinski-Harabasz index. In: IOP Conference Series: Materials Science and Engineering ). IOP Publishing (2019). Lai M, Demuru M, Hillebrand A, Fraschini M. A comparison between scalp- and source-reconstructed EEG networks. Sci Rep 8 , 12269 (2018). Kimes PK, Liu Y, Neil Hayes D, Marron JS. Statistical significance for hierarchical clustering. Biometrics 73 , 811-821 (2017). Helmer M , et al. On the stability of canonical correlation analysis and partial least squares with application to brain-behavior associations. Communications biology 7 , 217 (2024). Yizhar O, Fenno LE, Davidson TJ, Mogri M, Deisseroth K. Optogenetics in neural systems. Neuron 71 , 9-34 (2011). Canolty RT, Knight RT. The functional role of cross-frequency coupling. Trends in cognitive sciences 14 , 506-515 (2010). Petzschner FH, Garfinkel SN, Paulus MP, Koch C, Khalsa SS. Computational models of interoception and body regulation. Trends in neurosciences 44 , 63-76 (2021). Mulders D, Seymour B, Mouraux A, Mancini F. Confidence of probabilistic predictions modulates the cortical response to pain. Proceedings of the National Academy of Sciences 120 , e2212252120 (2023). Öztürk S, Devecioglu I, Güçlü B. Bayesian prediction of psychophysical detection responses from spike activity in the rat sensorimotor cortex. Journal of computational neuroscience 51 , 207-222 (2023). Uddin LQ, Kinnison J, Pessoa L, Anderson ML. Beyond the tripartite cognition-emotion-interoception model of the human insular cortex. Journal of cognitive neuroscience 26 , 16-27 (2014). Chen WG , et al. The Emerging Science of Interoception: Sensing, Integrating, Interpreting, and Regulating Signals within the Self. Trends Neurosci 44 , 3-16 (2021). Uddin LQ. Salience processing and insular cortical function and dysfunction. Nature reviews neuroscience 16 , 55-61 (2015). Van den Bergh O, Witthöft M, Petersen S, Brown RJ. Symptoms and the body: taking the inferential leap. Neuroscience & Biobehavioral Reviews 74 , 185-203 (2017). Zhang R, Deng H, Xiao X. The Insular Cortex: An Interface Between Sensation, Emotion and Cognition. Neuroscience Bulletin , 1-11 (2024). Akiki TJ, Jubeir J, Bertrand C, Tozzi L, Williams LM. Neural circuit basis of pathological anxiety. Nature reviews Neuroscience 26 , 5-22 (2025). Simmons WK, DeVille DC. Interoceptive contributions to healthy eating and obesity. Current opinion in psychology 17 , 106-112 (2017). Wiebking C, Bauer A, de Greck M, Duncan NW, Tempelmann C, Northoff G. Abnormal body perception and neural activity in the insula in depression: an fMRI study of the depressed "material me". The world journal of biological psychiatry : the official journal of the World Federation of Societies of Biological Psychiatry 11 , 538-549 (2010). Arminjon M, Preissmann D, Chmetz F, Duraku A, Ansermet F, Magistretti PJ. Embodied memory: unconscious smiling modulates emotional evaluation of episodic memories. Frontiers in psychology 6 , 650 (2015). Bai J, Kong S, Gomes C. Disentangled variational autoencoder based multi-label classification with covariance-aware multivariate probit model. In: Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence ) (2021). Barroso J, Branco P, Apkarian AV. The causal role of brain circuits in osteoarthritis pain. Nature Reviews Rheumatology , 1-14 (2025). Chang LJ, Yarkoni T, Khaw MW, Sanfey AG. Decoding the role of the insula in human cognition: functional parcellation and large-scale reverse inference. Cerebral cortex (New York, NY : 1991) 23 , 739-749 (2013). Li C , et al. Classification of schizophrenia spectrum disorder using machine learning and functional connectivity: reconsidering the clinical application. BMC Psychiatry 25 , 372 (2025). Liang Z, Ye W, Liu Q, Zhang L, Huang G, Zhou Y. NSSI-Net: A Multi-Concept GAN for Non-Suicidal Self-Injury Detection Using High-Dimensional EEG in a Semi-Supervised Framework. (2024). Krueger RF , et al. Progress in achieving quantitative classification of psychopathology. World Psychiatry 17 , 282-293 (2018). Thut G , et al. Guiding transcranial brain stimulation by EEG/MEG to interact with ongoing brain activity and associated functions: A position paper. Clinical neurophysiology : official journal of the International Federation of Clinical Neurophysiology 128 , 843-857 (2017). Zhao A, Fattahi D, Hu X. Physics-informed neural networks for physiological signal processing and modeling: a narrative review. Physiological measurement 46 , (2025). Aouedi O, Sacco A, Piamrat K, Marchetto G. Handling Privacy-Sensitive Medical Data With Federated Learning: Challenges and Future Directions. IEEE journal of biomedical and health informatics 27 , 790-803 (2023). Kop WJ, Toussaint A, Mols F, Löwe B. Somatic symptom disorder in the general population: Associations with medical status and health care utilization using the SSD-12. Gen Hosp Psychiatry 56 , 36-41 (2019). Bagby RM, Ryder AG, Schuller DR, Marshall MB. The Hamilton Depression Rating Scale: has the gold standard become a lead weight? American Journal of Psychiatry 161 , 2163-2177 (2004). Leentjens AF, Dujardin K, Marsh L, Richard IH, Starkstein SE, Martinez-Martin P. Anxiety rating scales in Parkinson's disease: a validation study of the Hamilton anxiety rating scale, the Beck anxiety inventory, and the hospital anxiety and depression scale. Movement Disorders 26 , 407-415 (2011). Petropoulakos K , et al. Validity and reliability of the Greek version of Pittsburgh Sleep Quality Index in chronic non-specific low back pain patients. In: Healthcare ). MDPI (2024). Delorme A, Makeig S. EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysis. Journal of neuroscience methods 134 , 9-21 (2004). Gramfort A , et al. MEG and EEG data analysis with MNE-Python. Frontiers in Neuroinformatics 7 , 267 (2013). Noack MM , et al. Gaussian processes for autonomous data acquisition at large-scale synchrotron and neutron facilities. Nature Reviews Physics 3 , 685-697 (2021). Bai J, Kong S, Gomes CP. Gaussian mixture variational autoencoder with contrastive learning for multi-label classification. In: international conference on machine learning ). PMLR (2022). Franco-Pereira AM, Nakas CT, Reiser B, Pardo MC. Inference on the overlap coefficient: the binormal approach and alternatives. Statistical methods in medical research 30 , 2672-2684 (2021). Liu Y, Hayes DN, Nobel A, Marron JS. Statistical significance of clustering for high-dimension, low–sample size data. Journal of the American Statistical Association 103 , 1281-1293 (2008). Zamani Esfahlani F, Byrge L, Tanner J, Sporns O, Kennedy DP, Betzel RF. Edge-centric analysis of time-varying functional brain networks with applications in autism spectrum disorder. Neuroimage 263 , 119591 (2022). Yeo BT , et al. The organization of the human cerebral cortex estimated by intrinsic functional connectivity. Journal of neurophysiology , (2011). Arzate-Mena JD , et al. Stationary EEG pattern relates to large-scale resting state networks–An EEG-fMRI study connecting brain networks across time-scales. NeuroImage 246 , 118763 (2022). Pu Y , et al. Variational autoencoder for deep learning of images, labels and captions. Advances in neural information processing systems 29 , (2016). Wu M, Goodman N. Multimodal generative models for scalable weakly-supervised learning. Advances in neural information processing systems 31 , (2018). Yi K, Wang Y, Ren K, Li D. Learning topology-agnostic EEG representations with geometry-aware modeling. In: Proceedings of the 37th International Conference on Neural Information Processing Systems ). Curran Associates Inc. (2023). Jiang WB, Zhao LM, Lu BL. Large Brain Model for Learning Generic Representations with Tremendous EEG Data in BCI. (2024). Tables Table 1. Performance Comparison of Representative Models under Model-Driven and Data-Driven Strategies Models Evaluation Criterion ACC AUC Sensitivity 1-Specificity VAE 59 0.77 0.79 0.76 0.78 MPVAE 37 0.78 0.80 0.78 0.80 POE-VAE 60 0.80 0.82 0.80 0.82 VAE-GM 59 0.80 0.82 0.81 0.82 MPVAE-GM 37 0.81 0.83 0.82 0.84 POE-VAE-GM 60 0.83 0.85 0.84 0.86 NSSINet 41 0.82 0.83 0.82 0.81 MMM 61 0.82 0.84 0.81 0.83 LaBraM 62 0.83 0.84 0.85 0.87 CVAE-GM 0.85 0.87 0.86 0.88 GM: Gaussian Mixture Additional Declarations There is NO Competing Interest. Not applicable Supplementary Files SupplementaryTables.zip Supplementary Tables: Brain region and cohort characteristics SupplementaryInformation.zip Supplementary Information: Model architecture, training configuration, and additional analyses Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7971302","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":537668132,"identity":"b818feb5-211b-4c51-892b-84d441abcb64","order_by":0,"name":"Nan Yan","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAzklEQVRIiWNgGAWjYBADOQZmUrUYA7UwNpCkJRGonEgt5uxnDzDz1NxJn9/O/PwBQ40dA/9sAjote/ISmHmOPcvdcJjNsIHhWDKDxJ0D+LUYHMgxYOZhO5y7gZkBqIXtAIOBRAIBLeffALX8O5wu38z+sYHhHzFabgBt4W07nMBwmMewgbGNKC1vDA7O7TtsuOEwT+GMxL5kHokbBB2WY/jgzbfD8vL9xzd8+PDNTo5/BgEtIHCIB8YCKubBoxABGH8QpWwUjIJRMApGLAAAJWlAtu8bCL4AAAAASUVORK5CYII=","orcid":"","institution":"Chinese Academy of Sciences, Shenzhen Institute of Advanced Technology (SIAT)","correspondingAuthor":true,"prefix":"","firstName":"Nan","middleName":"","lastName":"Yan","suffix":""},{"id":537668133,"identity":"98df5046-8248-40f9-a423-ff6a5d3026a4","order_by":1,"name":"Shuzhi Zhao","email":"","orcid":"","institution":"Chinese Academy of Sciences, Shenzhen Institute of Advanced Technology (SIAT)","correspondingAuthor":false,"prefix":"","firstName":"Shuzhi","middleName":"","lastName":"Zhao","suffix":""},{"id":537668134,"identity":"6cfe355f-0743-4545-b10c-696314395d49","order_by":2,"name":"Chongyuan Lian","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Chongyuan","middleName":"","lastName":"Lian","suffix":""},{"id":537668135,"identity":"7a99c9e8-8ef0-4c32-8a5a-d029185fee6f","order_by":3,"name":"Xue Shi","email":"","orcid":"","institution":"Shenzhen People’s Hospital","correspondingAuthor":false,"prefix":"","firstName":"Xue","middleName":"","lastName":"Shi","suffix":""},{"id":537668136,"identity":"e82cf55a-1d7b-4fbb-bdc1-9f11e5c8eba2","order_by":4,"name":"Ge Dang","email":"","orcid":"","institution":"Shenzhen People’s Hospital","correspondingAuthor":false,"prefix":"","firstName":"Ge","middleName":"","lastName":"Dang","suffix":""},{"id":537668137,"identity":"50e9bf20-6bf7-460f-8530-9f063a11e466","order_by":5,"name":"Zi’an Pei","email":"","orcid":"","institution":"Shenzhen People’s Hospital","correspondingAuthor":false,"prefix":"","firstName":"Zi’an","middleName":"","lastName":"Pei","suffix":""},{"id":537668138,"identity":"379b9a6f-df05-41a0-bb90-b3d02b93939d","order_by":6,"name":"Xiaoyong Lan","email":"","orcid":"","institution":"Shenzhen People’s Hospital","correspondingAuthor":false,"prefix":"","firstName":"Xiaoyong","middleName":"","lastName":"Lan","suffix":""},{"id":537668139,"identity":"0d672570-aac6-4e26-a372-e28d63b6c25d","order_by":7,"name":"Hanjun Liu","email":"","orcid":"","institution":"The First Affiliated Hospital, Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"Hanjun","middleName":"","lastName":"Liu","suffix":""},{"id":537668140,"identity":"da263da1-9ec5-465b-a391-f41c5caf1f86","order_by":8,"name":"Hiuching Hung","email":"","orcid":"","institution":"Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), Fremdsprachendidaktik","correspondingAuthor":false,"prefix":"","firstName":"Hiuching","middleName":"","lastName":"Hung","suffix":""},{"id":537668141,"identity":"a415b21d-1dd8-4a54-a41d-15182aa1f1cb","order_by":9,"name":"Dezhong Yao","email":"","orcid":"https://orcid.org/0000-0002-8042-879X","institution":"University of Electronic Science and Technology of China","correspondingAuthor":false,"prefix":"","firstName":"Dezhong","middleName":"","lastName":"Yao","suffix":""},{"id":537668142,"identity":"e395bfd7-0268-43ab-bddb-95fad8295f23","order_by":10,"name":"Lan Wang","email":"","orcid":"","institution":"Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Lan","middleName":"","lastName":"Wang","suffix":""},{"id":537668143,"identity":"8cacaa5d-d765-466b-bcfe-5f90ae3c733b","order_by":11,"name":"Xin Jiang","email":"","orcid":"","institution":"Department of Geriatrics, Shenzhen People's Hospital, The Second Clinical Medical College, Jinan University, The First Affiliated Hospital, Southern University of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Xin","middleName":"","lastName":"Jiang","suffix":""},{"id":537668144,"identity":"3bcf7bcd-aa0c-4cb9-a17b-487a31a38cc9","order_by":12,"name":"Yi Guo","email":"","orcid":"","institution":"Shenzhen People's Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yi","middleName":"","lastName":"Guo","suffix":""}],"badges":[],"createdAt":"2025-10-28 17:11:01","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7971302/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7971302/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":95224039,"identity":"2510d630-fbce-462e-9f95-67cbcbd834ae","added_by":"auto","created_at":"2025-11-05 16:23:14","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":231980,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eOverview of the CVAE-GM framework for multi-frequency EEG representation learning in SSD.\u003c/strong\u003e Raw SSD EEG signals are decomposed into low- and high-frequency bands and transformed into functional connectivity graphs. Within the Encoder (A), the Augmentation Generator (AugGen) produces frequency-specific perturbed graph views, which are projected into latent embeddings and optimized via contrastive loss. The latent space is regularized by a Gaussian Mixture Model prior (B) to capture multimodal characteristics. The Decoder (C) reconstructs EEG data from latent variables, while the clustering of latent representations reveals biologically meaningful subtypes, which are further mapped to cortical sources. The overall objective combines reconstruction, KL divergence, and contrastive losses, ensuring robust and interpretable subtype identification in SSD.\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-7971302/v1/d3d1384a3bb36d2ce6708968.png"},{"id":95224663,"identity":"ab8b5377-9681-4b9c-9370-0d50e127e68a","added_by":"auto","created_at":"2025-11-05 16:24:08","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":657758,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eContrastive VAE improves subtype separability and reveals neurobiological underpinnings in SSD. \u003c/strong\u003e(A) Comparison of standard VAE and contrastive VAE frameworks. While the VAE yields clusters with higher overlap (4.1%), incorporating multi-frequency contrastive loss markedly improves cluster separability (overlap reduced to 1.5%). (B) Cortical mapping of three SSD subtypes identified by the contrastive VAE. Distinct spatial patterns were observed: Cluster 1 was characterized by dominant alterations in the \u003cstrong\u003esomatomotor network (SMN)\u003c/strong\u003e, Cluster 2 in the \u003cstrong\u003ecentral executive network (CEN)\u003c/strong\u003e, and Cluster 3 in the \u003cstrong\u003elimbic network (LN)\u003c/strong\u003e. These subtype-specific network abnormalities highlight neurobiological heterogeneity underlying SSD symptom dimensions.\u003c/p\u003e","description":"","filename":"image2.png","url":"https://assets-eu.researchsquare.com/files/rs-7971302/v1/0c7a41f439742684906554ba.png"},{"id":95223953,"identity":"fd2a3a08-b23d-4222-9de0-7df14a41cca4","added_by":"auto","created_at":"2025-11-05 16:23:07","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":938626,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSubtype-specific cortical patterns and divergence analyses across SSD clusters. \u003c/strong\u003e(A–C) Visualization of cortical activation strength across individuals within each subtype (Cluster 1–3), sorted by distance to the cluster center (from center-close to center-distant). Consistent spatial patterns emerge within each cluster: Cluster 1 predominantly involves the \u003cstrong\u003esomatomotor network (SMN)\u003c/strong\u003e, Cluster 2 the \u003cstrong\u003ecentral executive network (CEN)\u003c/strong\u003e, and Cluster 3 the \u003cstrong\u003elimbic network (LN)\u003c/strong\u003e. Right panels display quantitative divergence analyses between subtypes using Kullback–Leibler (KL) divergence and Jensen–Shannon (JS) divergence, both showing significant differences (\u003cstrong\u003ep \u0026lt; 0.01\u003c/strong\u003e) across clusters, indicating robust subtype separability and distinct neural profiles.\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-7971302/v1/1c75390a085857459c0f26a5.png"},{"id":95088240,"identity":"d5486946-71c9-4a6e-85db-4c3c9e340329","added_by":"auto","created_at":"2025-11-04 07:47:30","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":348803,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAssociations between SSD subtypes, large-scale brain networks, and symptom dimensions. \u003c/strong\u003e(A) Radar plots showing network contribution patterns of three SSD subtypes across canonical large-scale networks (SMN, CEN, LN, DMN, DAN, VN). Subtype 1 is dominated by somatomotor network (SMN) alterations, Subtype 2 by central executive network (CEN), and Subtype 3 by limbic network (LN). Corresponding radar plots illustrate symptom associations across three clinical dimensions (somatic amplification, affective dysregulation, environmental stress sensitivity), with significant subtype-specific correlations (e.g., Subtype 1: r = 0.42, p \u0026lt; 0.001; Subtype 2: r = -0.38, p \u0026lt; 0.001; Subtype 3: r = 0.36, p = 0.002). (B) Replication analysis in an independent validation set confirms consistent subtype-to-network and subtype-to-symptom mappings, with significant correlations across all three subtypes (Subtype 1: r=0.40, p\u0026lt;0.001; Subtype 2: r=−0.37, p\u0026lt;0.001; Subtype 3: r=0.34, p=0.003).\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-7971302/v1/75287f7f9af3f9d025f03677.png"},{"id":95088245,"identity":"6e4709c8-3769-4822-ae02-8c101236aab7","added_by":"auto","created_at":"2025-11-04 07:47:30","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":651939,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSubtype-specific patterns in SSD compared with healthy controls. \u003c/strong\u003e(A) Cortical surface maps show altered connectivity profiles across three SSD subtypes (Cluster 1–3). Cluster 1 exhibited aberrant connectivity between the insula and the anterior cingulate cortex (ACC), premotor cortex (PMC), and precuneus. Cluster 2 showed abnormal insula–dorsolateral prefrontal cortex (DLPFC) and occipito-temporal (OT) connectivity. Cluster 3 demonstrated disrupted coupling between the insula and posterior cingulate cortex (PCC) and the thalamus. (B) Boxplots present quantitative comparisons between SSD subtypes and healthy controls (HC) for the key altered connections: insula–ACC in Cluster 1, insula–DLPFC in Cluster 2, and insula–thalamus in Cluster 3. All differences were statistically significant (p \u0026lt; 0.001).\u003c/p\u003e","description":"","filename":"image5.png","url":"https://assets-eu.researchsquare.com/files/rs-7971302/v1/d548f339b1c322732a7bc890.png"},{"id":95796990,"identity":"45058764-b8e1-4c8e-9941-f59c0d1ab4ff","added_by":"auto","created_at":"2025-11-13 08:00:06","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3973016,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7971302/v1/afa05bcc-bd96-4648-97e7-7c52e2d079f5.pdf"},{"id":95088239,"identity":"14a830ba-008f-4cb3-816c-9ea940de908e","added_by":"auto","created_at":"2025-11-04 07:47:30","extension":"zip","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":19620,"visible":true,"origin":"","legend":"Supplementary Tables: Brain region and cohort characteristics","description":"","filename":"SupplementaryTables.zip","url":"https://assets-eu.researchsquare.com/files/rs-7971302/v1/1efee0d68b5470aa37cbdabd.zip"},{"id":95088243,"identity":"9c921341-0c62-4a64-a862-5cdbd27a7ecc","added_by":"auto","created_at":"2025-11-04 07:47:30","extension":"zip","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":1368219,"visible":true,"origin":"","legend":"Supplementary Information: Model architecture, training configuration, and additional analyses","description":"","filename":"SupplementaryInformation.zip","url":"https://assets-eu.researchsquare.com/files/rs-7971302/v1/ccef86ccf619e1453bc1ebc3.zip"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.\nNot applicable","formattedTitle":"Multi-Frequency EEG Connectomics Uncovers Insula-Network Subtypes in Somatic Symptom Disorder","fulltext":[{"header":"Introduction","content":"\u003cp\u003eSomatic Symptom Disorder (SSD) imposes a substantial burden on healthcare systems, driven by a cycle of distressing physical symptoms and the excessive cognitive and emotional responses they provoke \u003csup\u003e1, 2\u003c/sup\u003e. Its high prevalence, affecting up to a third of patients in general hospitals \u003csup\u003e3\u003c/sup\u003e, underscores its substantial impact on public health. Although validated psychometric scales such as the Patient Health Questionnaire-15 (PHQ-15) \u003csup\u003e4\u003c/sup\u003e, the Somatic Symptom Disorder-B Criteria Scale (SSD-12) \u003csup\u003e5\u003c/sup\u003e, and the Neuro-11 \u003csup\u003e6\u003c/sup\u003e have been developed, the current categorical framework aggregates clinically heterogeneous presentations, which likely reflect distinct and overlapping neurobiological mechanisms \u003csup\u003e7, 8\u003c/sup\u003e. This heterogeneity results in more than 40% of patients failing to respond adequately to standard pharmacological and psychotherapeutic interventions \u003csup\u003e9\u003c/sup\u003e. These limitations underscore the need for a precision-medicine approach that integrates computational modeling with neurophysiological insights to enable personalized patient stratification and clinically interpretable subtyping.\u003c/p\u003e\n\u003cp\u003eSSD may be conceptualized as a disorder of hierarchical predictive coding, wherein aberrant precision weighting of interoceptive prediction errors amplifies innocuous bodily sensations into persistent and distressing symptoms \u003csup\u003e10\u003c/sup\u003e. Specifically, excessive precision of somatic priors prevents the appropriate updating of internal models, resulting in sustained prediction error signals that manifest as clinically unexplained symptoms \u003csup\u003e11, 12\u003c/sup\u003e. These computational disruptions are thought to engage mechanisms of central sensitization as a transdiagnostic process across multiple somatic domains. Central sensitization is hypothesized to be mediated by maladaptive, context-dependent functional alterations in brain regions that mediate interoceptive \u003csup\u003e10\u003c/sup\u003e, affective, and sensory processing, including the somatosensory cortex, prefrontal cortices, insula cortex, anterior cingulate cortex, thalamus, and brainstem. In this context, neuroimaging-based subtyping represents a promising avenue for identifying neurobiologically homogeneous patient subgroups, paving the way for the development of pathophysiological biomarkers and therapeutic interventions.\u003c/p\u003e\n\u003cp\u003eElectroencephalography (EEG) offers a powerful means to characterize the tempo-spatial dynamics of neural processes implicated in the clinical heterogeneity of SSD. Previous studies employing spectral power analysis \u003csup\u003e13, 14\u003c/sup\u003e, phase synchronization \u003csup\u003e14\u003c/sup\u003e, and functional connectivity \u003csup\u003e15\u003c/sup\u003e, along with machine learning architectures, have differentiated SSD patients from healthy controls. While previous supervised machine learning models have achieved high accuracy in distinguishing SSD patients from healthy controls, they fail to address the critical issue of patient-level heterogeneity, a key factor limiting treatment efficacy. This underscores a significant gap: the need for methods that can stratify patients rather than merely classify them. Self-supervised learning offers a promising solution by uncovering latent, data-driven structures without reliance on predefined diagnostic labels \u003csup\u003e16, 17\u003c/sup\u003e. Therefore, the objective of this study was to develop and validate a computational framework using a contrastive variational autoencoder (CVAE) applied to large-scale EEG connectomics.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eHowever, applying these subtyping methods faces significant challenges \u003csup\u003e18\u003c/sup\u003e. Current models often analyze single-frequency bands, ignoring crucial cross-frequency dynamics implicated in SSD, a critical aspect of the neurophysiological principle underlying complex cognitive processes. They also focus on group-level contrasts, failing to capture critical within-group heterogeneity. Finally, their inherent lack of interpretability acts as a major barrier, preventing the translation of model features into mechanistic insights about symptom drivers. Although post-hoc attribution methods (e.g., SHAP, Integrated Gradients) can identify salient features, they fall short of elucidating the causal or theoretical relationships between neural dynamics and symptom dimensions.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo address these limitations, we introduce a computational framework that identifies neurophysiological subtypes of SSD by integrating multi-frequency EEG connectivity with self-supervised learning. Our model leverages multi-band cross-attention to capture hierarchical neural dynamics, a key aspect of neural multiplexing. Contrastive learning then identifies patient subgroups as deviations from a normative neural manifold, while a Gaussian mixture model accommodates the disorder\u0026apos;s profound heterogeneity. The framework successfully identifies robust, reproducible subtypes whose distinct connectivity patterns are significantly associated with specific symptom dimensions. This integrative approach represents a paradigm shift from phenomenological classification toward a mechanistically-informed psychiatric nosology, offering a path to enhanced diagnostic precision and the design of individualized interventions.\u003c/p\u003e\n\u003cp\u003e[\u003cstrong\u003eInsert Figure 1 near here\u003c/strong\u003e]\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cstrong\u003eCan the integration of cross-frequency EEG connectome analysis with contrastive learning methods identify neurobiologically distinct subtypes within SSD?\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe investigated whether our conditional variational autoencoder with Gaussian mixture modeling (CVAE-GM) could identify biologically meaningful subtypes within the large SSD electroencephalography dataset (n = 1,419). To determine the optimal cluster number, we applied the Calinski-Harabasz index \u003csup\u003e19\u003c/sup\u003e alongside distributional analysis of within-cluster scalp network strength. While this index suggested either two or three clusters (Supplementary Fig. 1A), distributional analysis \u003csup\u003e20\u003c/sup\u003e revealed that the two-cluster solution produced multimodal distributions within clusters, indicating residual heterogeneity (Supplementary Fig. 1B). Conversely, the three-cluster solution yielded unimodal, well-concentrated distributions (Supplementary Fig. 1C), leading us to select k = 3 for subsequent analyses.\u003c/p\u003e\n\u003cp\u003e[\u003cstrong\u003eInsert Figure 2 near here\u003c/strong\u003e]\u003c/p\u003e\n\u003cp\u003eThe CVAE-GM model identified three distinct subtypes with minimal overlap (mean = 1.5%), achieving a 62.5% improvement in cluster separability over standard variational autoencoders (4.1% overlap). Features from theta and beta frequency bands provided the highest discriminatory power, highlighting the importance of frequency-specific neural dynamics in subtype differentiation. The three subtypes were characterized by distinct network dominance patterns: sensorimotor network (SMN)-dominant, central executive network (CEN)-dominant, and limbic network (LN)-dominant (Fig. 2), each of which is defined by the distinct functional activation pattern represented by its cluster centroid.\u003c/p\u003e\n\u003cp\u003eTo assess intra-subtype heterogeneity, we examined continuous variation in connectivity patterns as a function of distance from each cluster\u0026apos;s centroid. This analysis is predicated on the principle that a continuous representation of an individual\u0026apos;s similarity to a subtype centroid provides a finer-grained characterization of inter-individual differences than categorical assignment alone (See Supplementary\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eAppendix). \u0026nbsp;Individual connectivity maps showed progressive divergence from centroid prototypes with increasing distance across all three subtypes (Fig. 3A\u0026ndash;C). Kullback-Leibler and Jensen-Shannon divergence analyses confirmed substantial between-subtype differentiation with low cross-modal concordance (Fig. 3D).\u003c/p\u003e\n\u003cp\u003e[\u003cstrong\u003eInsert Figure 3 near here\u003c/strong\u003e]\u003c/p\u003e\n\u003cp\u003eInternal validation confirmed subtype robustness across multiple approaches, as proposed by \u003csup\u003e18\u003c/sup\u003e. SigClust analysis \u003csup\u003e21\u003c/sup\u003e established statistical significance for all three subtypes (p_FDR \u0026lt; 0.001). Consistency metrics demonstrated excellent clustering performance (V-measure: 0.87; adjusted Rand index: 0.88; normalized mutual information: 0.86; silhouette coefficient: 0.84) (Supplementary Fig. 1c).\u0026nbsp;Overall, these analyses suggest that the identified neuro-subtypes are internally valid and likely reflect the heterogeneity of the EEG data. Note that because of the limitations of the testing methods, these analyses provide evidence but do not guarantee the separability of the identified subtypes. In the absence of ground truth, internal validation of clustering results remains challenging, thus necessitating external validation using metrics that were not employed in the clustering analysis.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDo neurobiologically distinct subtypes exhibit specific clinical profiles?\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe Canonical Correlation Analysis (CCA) \u003csup\u003e22\u003c/sup\u003e was employed to assess the multivariate relationship between subtype-specific functional connectivity density (FCD) patterns and the clinical dimensions of the Neuro-11 questionnaire. In this analysis, we found significant specific mappings (See Figure 4): The SMN-dominant subtype was characterized by elevated FCD within the SMN, which was significantly and positively correlated with somatic symptom scores on the Neuro-11 scale (r = 0.42, p \u0026lt; 0.001). Similarly, the CEN-domain subtype exhibited increased FCD within the CEN, which was negatively correlated with scores on the negative life events dimension (r =\u0026ndash;0.38, p \u0026lt; 0.001). Finally, the LN-domain subtype exhibited heightened FCD within the LN, which was positively correlated with emotion dimension scores (r = 0.36, p = 0.002). In contrast, no other networks showed significant clinical associations (all p \u0026gt; 0.05).\u003c/p\u003e\n\u003cp\u003e[\u003cstrong\u003eInsert Figure 4 near here\u003c/strong\u003e]\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eShared and distinct neural substrates across SSD subtypes\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo validate model robustness, we incorporated age- and sex-matched healthy controls (n = 112). The CVAE-GM model maintained effectiveness, identifying four clusters with 8.5% mean overlap. Healthy controls formed a distinct cluster characterized by default mode network dominance in theta and beta bands, consistent with established neurophysiology. The three patient subtypes remained stable (Supplementary Fig. 2).\u003c/p\u003e\n\u003cp\u003eFDR-Group-level comparisons between each subtype and controls revealed distinct neural substrates. The SMN-dominant subtype exhibited hyperactivation in the anterior cingulate cortex (ACC), ventrolateral prefrontal cortex (vlPFC), insula, and premotor cortex, suggesting overactive interoceptive-sensorimotor integration. The CEN-dominant subtype showed aberrant activity in the dorsolateral prefrontal cortex (DLPFC), precuneus, and insula, reflecting dysregulated inhibitory control. The LN-dominant subtype displayed hyperactivity in the thalamus, posterior cingulate cortex (PCC), and occipitotemporal junction, indicating limbic-autonomic dysfunction.\u003c/p\u003e\n\u003cp\u003e[\u003cstrong\u003eInsert Figure 5 near here\u003c/strong\u003e]\u003c/p\u003e\n\u003cp\u003eNotably, insula dysfunction was observed across all subtypes, suggesting a core pathophysiological hub. Connectivity analyses revealed subtype-specific insula coupling patterns: altered insula\u0026ndash;anterior cingulate connectivity (SMN-dominant), disrupted insula\u0026ndash;dorsolateral prefrontal connectivity (CEN-dominant), and abnormal insula\u0026ndash;thalamus connectivity (LN-dominant) (Fig. 5B). These aberrant connectivity pathways reinforce the notion that dysfunctional insula interactions with specific brain regions may represent a central neurobiological mechanism underlying different symptomatic manifestations of SSD.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eIndependent validation confirms subtype reproducibility\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eOur analyses revealed a critical dissociation: while distinct data modalities independently identified robust subtypes with convergent clinical profiles. A central question, therefore, is whether these findings are specific to the cohort under investigation or reflect broader patterns observable across diverse populations.\u0026nbsp;We validated our findings using an independent SSD dataset (n = 530). Participants were assigned subtype labels based on predominant clinical dimensions (somatic, negative emotion, or adverse events) from Neuro-11 assessments. A comparative analysis of state-of-the-art self-supervised learning methods demonstrated that all approaches achieved high classification accuracy, confirming the underlying heterogeneity of SSDs. Our CVAE-GM framework achieved superior performance (accuracy: 0.85; AUC: 0.87; sensitivity: 0.86; 1-specificity: 0.88) (Table 1), validating both the existence of distinct neurophysiological subtypes and the effectiveness of our methodology in capturing this heterogeneity.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e[\u003cstrong\u003eInsert Table 1 near here\u003c/strong\u003e]\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we propose a framework for stratifying individuals with SSD based on latent neural signatures associated with symptom heterogeneity. By integrating contrastive variational autoencoders, Gaussian mixture modeling, and multi-band EEG functional connectivity, our model identified three distinct SSD subtypes characterized by distinct patterns of neural dysregulation that corresponded to specific symptom dimensions measured by the Neuro-11 scale. These results move beyond traditional diagnostic categories by establishing a mechanistic basis for data-driven subtyping, offering a concrete step toward precision psychiatry in functional somatic syndromes.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eNeurobiological mechanisms underlying SSD heterogeneity.\u003c/strong\u003e Our findings indicate that symptom heterogeneity in SSD is underpinned by distinct patterns of CFC dysfunction across large-scale networks. In particular, the SMN-dominant subtype exhibited hyperconnectivity within sensorimotor circuits (P\u0026lt;0.001), characterized by \u0026theta;-\u0026beta; CFC enhancement between\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003ethe insula and key regions involved in sensorimotor integration, including anterior cingulate cortex (ACC, modulation index: 0.65\u0026plusmn;0.05 vs 0.45\u0026plusmn;0.07), premotor cortex (PMC, 0.60\u0026plusmn;0.06 vs 0.44\u0026plusmn;0.07), and precuneus (0.60\u0026plusmn;0.05 vs 0.43\u0026plusmn;0.06; all \u003cem\u003eP\u003c/em\u003e \u0026lt; 0.001). This CFC enhancement was significantly correlated with somatic symptom severity (r = 0.46, P \u0026lt; 0.001). This pattern is consistent with disrupted GABA-mediated inhibitory control within sensorimotor circuits \u003csup\u003e23\u003c/sup\u003e, compromising the hierarchical predictive coding processing required for accurate body-state estimation. Aberrant \u0026theta;-\u0026beta; coupling disrupts this balance, leading to misaligned prediction-error signaling whereby innocuous interoceptive stimuli are misinterpreted as salient bodily threats \u003csup\u003e24\u003c/sup\u003e. This is consistent with models of dysfunctional Bayesian inference, wherein overly precise prediction errors amplify normal bodily signals into exaggerated somatic experiences \u003csup\u003e25\u003c/sup\u003e.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFunctionally, the observed insula-ACC-PMC-precuneus hyperconnectivity implicates a dysfunctional network responsible for integrating interoceptive awareness (insula), emotional salience (ACC), motor intentions (PMC), and self-referential bodily representation (precuneus). Such dysregulation may contribute to a fragmented body schema and an exaggerated sense of symptom authenticity. Importantly, this neurophysiological signature helps explain why patients in the SMN-dominant subtype experience their symptoms as subjectively real despite the absence of identifiable peripheral pathology. Rather than reflecting mere emotional amplification or executive misattribution, the findings point to a fundamental miscalibration of internal sensorimotor models, wherein aberrant oscillatory dynamics generate persistent and biologically encoded prediction errors \u003csup\u003e26, 27\u003c/sup\u003e.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe CEN-dominant subtype exhibited increased connectivity density within fronto-parietal circuits (P \u0026lt; 0.001) and enhanced dlPFC-anterior insula coupling (0.63\u0026plusmn;0.05 vs 0.46\u0026plusmn;0.06, P \u0026lt; 0.001). These network alterations were inversely correlated with the negative event reactivity of Neuro-11 (r=-0.38, P \u0026lt; 0.001). This \u0026ldquo;hyperregulation\u0026rdquo; state represents a pathological configuration in which excessive top-down cognitive control disrupts adaptive interoceptive processing \u003csup\u003e10\u003c/sup\u003e, producing a form of cognitive override that amplifies, rather than attenuates, somatic symptoms through hypervigilant monitoring and catastrophic misinterpretation of bodily sensations \u003csup\u003e28, 29, 30\u003c/sup\u003e. The observed neural hyperactivation is therefore paradoxical: it represents an energetically expensive, dysfunctional process where cognitive engagement with interoceptive signals perpetuates symptomatology instead of providing regulatory control \u003csup\u003e31\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eThe LN-dominant subtype exhibited significantly increased functional connectivity within limbic circuits compared to healthy controls (\u003cem\u003eP\u003c/em\u003e \u0026lt; 0.001), with particularly robust enhancement along the thalamo-insula pathway (connectivity strength: 0.61\u0026plusmn;0.06 vs 0.44\u0026plusmn;0.06, \u003cem\u003eP\u003c/em\u003e \u0026lt; 0.001). This hyperconnectivity was positively correlated with emotional symptom severity of the Neuro-11 scale (r = 0.36, p = 0.002), indicating a convergent limbic-somatic pathophysiology \u003csup\u003e30, 32, 33\u003c/sup\u003e. Thalamo-insula hyperconnectivity appears to drive a limbic-mediated pathway of symptom amplification \u003csup\u003e34\u003c/sup\u003e. Within this circuit, excessive thalamic input to the posterior insula likely imbues normal bodily sensations with aberrant threat salience \u003csup\u003e35\u003c/sup\u003e. This mechanism stands in stark contrast to the top-down cognitive hypercontrolled of the CEN-dominant subtype, explaining a phenotype in which patients recognize the irrationality of their symptoms yet cannot override the powerful, affectively charged interoceptive experience \u003csup\u003e36\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eThe insula functions as a pathological convergence hub driving subtype-specific symptom generation in somatic symptom disorders.\u0026nbsp;\u003c/strong\u003eAcross all three SSD subtypes, we observed consistent abnormalities in insular connectivity, suggesting this region as a central hub in the pathophysiology of SSD. This convergence aligns with the insula\u0026rsquo;s unique anatomical and functional properties: it receives direct ascending interoceptive input via lamina I spinothalamic pathways and contains Von Economo neurons that support rapid interoceptive\u0026ndash;cognitive integration \u003csup\u003e24, 37\u003c/sup\u003e. These features position the insula as a key cortical interface through which bodily sensations acquire conscious representation and affective significance \u003csup\u003e10\u003c/sup\u003e. Notably, the clinical heterogeneity of SSD appears to arise from subregion-specific patterns of insular dysconnectivity. In SMN-dominant patients, posterior insular hyperconnectivity with somatosensory cortices is associated with heightened bodily awareness and amplified sensorimotor salience. In LN-dominant cases, enhanced anterior insula\u0026ndash;amygdala coupling correlates with emotional symptom severity, implicating dysregulated limbic integration of interoceptive signals. In CEN-dominant subtypes, increased connectivity between the dorsal anterior insula and dlPFC reflects maladaptive cognitive monitoring of internal bodily states \u003csup\u003e33, 38\u003c/sup\u003e. These findings support a tripartite model of insular dysfunction that reconciles SSD heterogeneity within a unified neurobiological framework: hyperactive posterior insula processing amplifies somatic sensations; dysregulated mid-insula integration perturbs salience attribution; and aberrant anterior insula coupling with executive or limbic circuits determines whether symptoms manifest as cognitive hypervigilance, emotional dysregulation, or sensorimotor distortion \u003csup\u003e39\u003c/sup\u003e. The insula thus emerges as both the proximal generator of symptom experience and a promising therapeutic target, where subtype-specific modulation of insular circuitry may offer a promising pathway for restoring adaptive interoceptive processing \u003csup\u003e10\u003c/sup\u003e and reducing symptom burden across heterogeneous SSD profiles.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eNeurobiologically-informed EEG subtyping transcends symptom-based taxonomies to enable precision psychiatry in somatic symptom disorders.\u0026nbsp;\u003c/strong\u003eCurrent psychiatric classifications yield treatment failure rates of 30-40% in SSD, reflecting their reliance on symptom-based categorization that ignores underlying pathophysiology \u003csup\u003e40\u003c/sup\u003e. While standardized clinical instruments such as the PHQ-15 and SCL-90 achieve moderate diagnostic accuracy, their dependence on subjective self-report limits their capacity to capture the multidimensional neurophysiological heterogeneity that drives symptom expression \u003csup\u003e3, 8, 41\u003c/sup\u003e. The Hierarchical Taxonomy of Psychopathology (HiTOP) represents a significant conceptual advance by modeling psychopathology along dimensional spectra, and tools such as the PID-5 have achieved improved predictive validity (AUC = 0.81-0.84) across both categorical and dimensional assessments \u003csup\u003e42\u003c/sup\u003e. However, HiTOP remains fundamentally symptom-centric, offering limited insight into the neural mechanisms that are essential for biologically informed, mechanism-targeted interventions. Our proposed EEG-based machine learning framework addresses these limitations by quantifying the neural substrates underlying SSD heterogeneity. The CVAE-GM architecture with multi-frequency contrastive learning achieved superior classification performance (balanced accuracy = 0.85, AUC = 0.87) compared to symptom-based approaches, while identifying three distinct neurofunctional subtypes with reproducible connectivity signatures. Unlike conventional diagnostic systems that describe phenomenology, therefore, our framework enables the direct identification of mechanistic subtypes to enhance diagnostic precision and inform individualized therapeutic strategies for SSD.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethodological advances and comparative performance.\u0026nbsp;\u003c/strong\u003eOur CVAE-GM framework exhibited substantial methodological advances over conventional approaches, achieving a 61.1% improvement in subtype stability (cluster separability silhouette coefficient: 0.87 vs 0.54 for Gaussian mixture models, P \u0026lt; 0.001) and enhanced discriminative power (AUC: 0.87, 95% CI: 0.85-0.93 vs 0.77, 95% CI: 0.71-0.81 for VAE). A central advantage of our framework lies in its contrastive learning architecture, which captures the distributed and frequency-specific nature of neural representations via multiband latent embeddings. This enables the identification of population-level oscillatory patterns that elude traditional univariate or frequency-agnostic models. Performance gains were further driven by a novel coupling consistency loss function that imposes neurobiologically informed constraints on cross-frequency interactions, mirroring neural ensemble synchronization mechanisms observed in healthy cortical circuits. To our knowledge, this represents the first participant-level clustering method that utilizes multiband EEG features in conjunction with Gaussian mixture contrastive learning to derive biologically informed subtypes. Importantly, it addresses a fundamental limitation of existing models that treat EEG as static connectivity matrices, instead modeling the dynamic and oscillatory nature of neural communication that underlies complex psychiatric phenotypes. The resulting subtypes are not only statistically robust but also neurobiologically plausible, exhibiting connectivity signatures consistent with established principles of cortical organization and cross-frequency coupling.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLimitations and future directions.\u003c/strong\u003e Several limitations of this study should be acknowledged. First, while our resting-state EEG approach captures static functional connectivity patterns, it may not fully reflect dynamic or context-dependent neural dysfunctions. Future studies should incorporate task-based paradigms targeting interoceptive attention, emotional regulation, and executive control, as well as perturbational techniques such as TMS\u0026ndash;EEG or neurofeedback, to establish causal relationships between network dynamics and behavioral phenotypes \u003csup\u003e43\u003c/sup\u003e. Second, the cross-sectional design and demographically homogenous sample (n = 150, exclusively Chinese) limit generalizability. Longitudinal studies tracking connectivity trajectories over time, along with multi-site replications in ethnically diverse populations, are essential to evaluate cultural influences and ensure cross-populational robustness. Third, combining EEG subtypes with high-resolution fMRI, polygenic risk scores, and peripheral biomarkers could yield comprehensive neurobiological profiles. And emerging computational advances, including physics-informed neural networks incorporating biophysical constraints and federated learning enabling multi-site validation, could offer promising avenues for developing diagnostic and therapeutic algorithms that transform SSD from heterogeneous diagnostic categories into distinct, biologically-defined conditions \u003csup\u003e44, 45\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions and translational implications.\u003c/strong\u003e We present a novel framework that elucidates the neural underpinnings of symptom dimensions in SSD using resting-state EEG, identifying three robust, neurobiologically-distinct subtypes. The clinical validity of these subtypes is supported by their strong correspondence with Neuro-11 symptom dimensions and distinct spectral patterns, particularly in cross-frequency coupling dynamics. These findings establish a data-driven, mechanistically grounded approach for stratifying SSD, providing significant insights into translating neurofunctional subtypes into personalized diagnostic and therapeutic strategies.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cstrong\u003eParticipants\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA total of 1,949 individuals diagnosed with somatic symptom disorder (SSD) and 112 healthy controls were recruited across multiple clinical departments at Shenzhen People\u0026apos;s Hospital. The SSD cohort was stratified into an internal training dataset (n = 1,419; mean age 48.11 \u0026plusmn; 15.70 years) recruited from Neurology departments, and an external validation dataset (n = 530; mean age 52.23 \u0026plusmn; 16.11 years) recruited from Cardiology, Gastroenterology, Endocrinology, and Traditional Chinese Medicine departments. Healthy controls (n = 112; mean age 42.03 \u0026plusmn; 13.72 years) were age-matched to the patient cohorts. To enhance ecological validity, both inpatient and outpatient populations were included. All participants provided informed consent before their participation, and procedures were approved by the institutions\u0026rsquo; Institutional Review Boards.\u003c/p\u003e\n\u003cp\u003eAll patients met the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) \u003csup\u003e46\u003c/sup\u003e criteria for SSD, characterized by persistent somatic symptoms accompanied by excessive health-related thoughts, feelings, and behaviors. Diagnostic assessments were conducted by experienced psychiatrists using a battery of standardized clinical instruments, including the Hamilton Depression Rating Scale (HAMD) \u003csup\u003e47\u003c/sup\u003e, the Hamilton Anxiety Rating Scale (HAMA) \u003csup\u003e48\u003c/sup\u003e, the Pittsburgh Sleep Quality Index (PSQI) \u003csup\u003e49\u003c/sup\u003e, and the Neuro-11 questionnaire \u003csup\u003e6\u003c/sup\u003e. Crucially, the final diagnosis was based on comprehensive clinical judgment rather than predetermined scale cutoff scores.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eExclusion criteria included: a history of severe psychiatric disorders (e.g., schizophrenia, bipolar disorder, major depressive disorder); current use of psychotropic medications or neuromodulation treatments (e.g., repetitive transcranial magnetic stimulation); significant cognitive impairment (e.g., dementia), language barriers, or impaired consciousness; substance abuse, suicidal risk, or recent cranial trauma or neurological disorders (e.g., epilepsy, Parkinson\u0026rsquo;s disease, amyotrophic lateral sclerosis); pregnancy or lactation; and any other medical or ethical considerations deemed unsuitable for study participation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEEG Acquisition and Signal Processing\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eResting-state EEG data (8 min eyes-open, 8 min eyes-closed) were acquired from 64 channels in a 10-10 configuration using a BrainAmp DC amplifier (Brain Products GmbH) at a sampling rate of 1,000 Hz. Impedances were kept below 5 k\u0026Omega;, with FCz serving as the reference and AFz as the ground. The raw data were processed using a pipeline incorporating EEGLAB \u003csup\u003e50\u003c/sup\u003e and MNE-Python \u003csup\u003e51\u003c/sup\u003e. The pipeline included 50 Hz notch and 1-45 Hz band-pass filtering, manual rejection of non-stereotyped artifacts, bad channel interpolation, and ICA-based removal of ocular and cardiac artifacts. The cleaned data were subsequently re-referenced to the common average. From these signals, we extracted the power spectral density (PSD) via FFT for six frequency bands: delta (1-4 Hz), slow theta (4-6 Hz), fast theta (6-8 Hz), alpha (8-13 Hz), beta (13-30 Hz), and gamma (30-45 Hz). Finally, cortical source activity was estimated using a standard forward and inverse modeling approach.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eIdentification of SSD Subtypes via a CVAE-GM Model\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe propose a novel deep learning model, the CVAE-GM, to identify distinct subtypes of SSD from neurophysiological data. This model combines a Conditional Variational Autoencoder (CVAE) with a Gaussian Mixture (GM) model in its latent space.\u0026nbsp;This architecture addresses the limitations of standard variational autoencoders, which assume unimodal latent distributions and inadequately capture heterogeneous clinical populations. The model integrates three key components: (1) A multi-band contrastive learning module that enforces feature invariance across frequency ranges while maintaining individual-specific pathophysiological signatures; (2) A Gaussian mixture latent space that replaces the standard unimodal Gaussian prior with a multimodal distribution, enabling direct modeling of distinct neural phenotypes; (3) An integrated training framework that embeds contrastive learning within the CVAE-GM optimization loop, yielding a latent space optimized for generation, clustering, and cross-frequency dynamics capture. \u0026nbsp;(see Supplementary Materials for architectural and training details).\u003c/p\u003e\n\u003cp\u003e1) \u003cstrong\u003eMulti-band Contrastive Learning Module\u003c/strong\u003e. To generate robust embedding feature representations, our model employs a contrastive learning objective across frequency bands. This module learns an embedding space where features from low-frequency (delta, theta) and high-frequency (beta, gamma) bands of the same individual are pulled closer, while those from different individuals are pushed apart. This process enforces the learning of features that are invariant across frequency ranges but specific to an individual\u0026apos;s pathophysiology, thereby enhancing subtype separation.\u003c/p\u003e\n\u003cp\u003e2) \u003cstrong\u003eGaussian Mixture Latent Space.\u003c/strong\u003e Standard VAEs typically assume a unimodal Gaussian prior \u003csup\u003e52, 53\u003c/sup\u003e, an oversimplification for heterogeneous clinical populations like SSD. To address this, our model\u0026apos;s latent space is regularized by a Gaussian Mixture Model prior. This allows the encoder to map individuals to distinct clusters within the latent space, directly modeling the multimodal distribution of neural phenotypes and improving the accuracy of subtype assignment.\u003c/p\u003e\n\u003cp\u003e3) \u003cstrong\u003eIntegrated Architecture for Subtyping.\u0026nbsp;\u003c/strong\u003eThe final architecture embeds the contrastive learning objective within the CVAE-GM training loop. This unified approach ensures that the model learns a latent space that is not only well-structured for generation (via the VAE) and clustering (via the GMM) but also enriched with features that capture cross-frequency dynamics (via contrastive learning). The result is a robust framework for data-driven subtyping that yields interpretable neural clusters to inform precision medicine.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSubtype validation framework using the comprehensive metrics\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eSubtype Stability Analysis\u003c/em\u003e\u003c/strong\u003e\u003cstrong\u003e.\u003c/strong\u003e A stable clustering solution should be robust to small perturbations of the dataset, such as removing a subset of subjects from the clustering procedure. We verified the stability of the identified subtypes by running the CVAE-GM model 100 times with different random initializations. The consistency of the resulting subject assignments was measured by calculating the average cluster overlap, which is a quantitative metric used to assess the stability and robustness of a clustering solution. We used the\u003cstrong\u003e\u0026nbsp;Overlap Coefficient (OVL)\u0026nbsp;\u003c/strong\u003eto quantify the overlap between subtype connectivity distributions and the standard normal distribution, thereby detecting residual heterogeneity or multimodality within subtypes. Estimation followed the parametric binormal and nonparametric kernel approach described by Franco‐Pereira et al. \u003csup\u003e54\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eSubtype validity and robustness Analysis.\u003c/em\u003e\u003c/strong\u003e To comprehensively evaluate the validity and robustness of the identified SSD subtypes, we employed a suite of both external and internal validation metrics. Individual-centroid divergence was measured using Kullback-Leibler (KL) and Jensen-Shannon divergences (JSD), capturing directional bias and overall similarity, respectively. For each subject, KL and JSD were calculated against the three centroids, and group-level differences were assessed using nonparametric rank tests with FDR correction.\u003c/p\u003e\n\u003cp\u003eExternal validation was performed by comparing our model\u0026apos;s subtype assignments to core dimensions of Neuro-11. We quantified the concordance using three distinct metrics: the Normalized Mutual Information (NMI), which measures the mutual dependence between partitions; the Adjusted Rand Index (ARI), which calculates similarity while correcting for chance; and the V-measure, which balances cluster homogeneity and completeness. For internal validation, which assesses the intrinsic structure of the data without external labels, we calculated the average silhouette coefficient. This metric quantifies the degree of cluster definition by measuring the ratio of inter-cluster separation to intra-cluster cohesion. A high degree of convergence across these diverse metrics would provide strong evidence for a meaningful and reliable subtyping solution.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eSignificance of the subtypes.\u003c/em\u003e\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eTo formally assess the statistical significance of the identified subtype structure, we employed the Significance of Clustering (SigClust) method \u003csup\u003e55\u003c/sup\u003e. This method formally evaluates the null hypothesis that the observed data originates from a single multivariate Gaussian distribution, as opposed to a Gaussian mixture. Through Monte Carlo simulation, SigClust calculates a p-value for this null hypothesis. A significant p-value indicates that the single-Gaussian model is a poor fit, thus providing statistical support for the presence of a genuine cluster structure in the data. Additionally, we constructed a continuous similarity gradient within each subtype. Specifically, each individual\u0026rsquo;s functional connectivity pattern was compared with the subtype centroid (using correlations of connectivity density maps), and subjects were ordered by similarity to form a gradient from \u0026ldquo;center\u0026rdquo; to \u0026ldquo;periphery.\u0026rdquo; This approach, adapted from prior work on continuous subtyping \u003csup\u003e15, 56\u003c/sup\u003e, provides a finer characterization of intra-subtype heterogeneity.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDetecting the heterogeneous patterns of functional brain activity among subtypes\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo elucidate the neurofunctional significance of the SSD subtypes identified by the CVAE-GM model, we performed a series of functional activity and Graph-based analyses on the source-reconstructed EEG data. First, the minimum norm estimate (MNE) method was applied to reconstruct resting-state EEG signals in source space, mapping electrophysiological activity to the cortical surface to obtain region-level time-series activity patterns. We applied the seven-network Yeo parcellation \u003csup\u003e57\u003c/sup\u003e to investigate alignment between CVAE-GM patterns and established functional architectures. Neural dynamics were characterized through phase-space reconstruction and attractor analysis \u003csup\u003e58\u003c/sup\u003e to detect transitional activation patterns indicative of unstable neural states.\u003c/p\u003e\n\u003cp\u003eGraph-theoretical analysis computed intra-network connection density for each functional network, defined as the ratio of existing to possible connections within networks. This quantified network integration and local specialization across subtypes. Group comparisons with healthy controls employed two-sample t-tests with false discovery rate correction (p \u0026lt; 0.05) to identify subtype-specific neural alterations.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAssociation Analysis between Subtypes and Clinical Assessments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe used Canonical Correlation Analysis (CCA) \u003csup\u003e22\u003c/sup\u003e to map the CVAE-GM-derived neural subtypes onto clinical symptom dimensions. The analysis tested for significant multivariate correlations between the neural latent features and scores from three domains of the Neuro-11 questionnaire \u003csup\u003e6\u003c/sup\u003e: somatic symptoms, negative affect, and adverse life events. By interpreting the significant canonical loadings, we identified the specific clinical profiles most characteristic of each neural subtype, thereby providing insight into the neural basis of clinical heterogeneity in SSD.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eExternal Validation on an Independent Dataset.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo assess the generalizability and robustness of the proposed CVAE-GM model, we conducted a comprehensive evaluation on a large, independent EEG dataset of individuals with SSD. The model was trained exclusively on the source dataset and tested on the independent cohort. Ground-truth labels for classification were assigned based on the dominant symptom dimension identified by the Neuro-11 questionnaire.\u003c/p\u003e\n\u003cp\u003eWe established a unified comparative framework to benchmark CVAE-GM against nine representative self-supervised learning models, including VAE \u003csup\u003e59\u003c/sup\u003e, MPVAE \u003csup\u003e37\u003c/sup\u003e, POE-VAE \u003csup\u003e60\u003c/sup\u003e, and their GM-integrated counterparts, as well as three state-of-the-art models (NSSINet \u003csup\u003e41\u003c/sup\u003e, MMM \u003csup\u003e61\u003c/sup\u003e, LaBraM \u003csup\u003e62\u003c/sup\u003e). To ensure a fair and reproducible comparison, all models were trained using a five-fold cross-validation scheme with identical input features (multi-band functional connectivity), optimizer, learning rate, batch size, and a maximum of epochs with an early stopping policy. All experiments were executed within a standardized hardware and software environment. Model performance was quantified using Accuracy (ACC), Area Under the Curve (AUC), Sensitivity, and Specificity.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eCode availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll custom source code developed for this study, including the implementation of the conditional variational autoencoder with Gaussian mixture prior (CVAE-GM), is publicly available at https://github.com/zhaoshuhzi/CVAE-GM under the MIT License. The repository includes scripts, model definitions, and documentation necessary to reproduce the analyses and results presented in this work.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe thank all members of the Yan and Guo laboratories for their valuable discussions and technical assistance throughout this project. We are also grateful to Prof. Dezhong Yao and Prof. Lan Wang for their insightful comments on the analysis and interpretation of the data. This work was supported by the National Natural Science Foundation of China (U23B2018, 62271477, 82371471), the Shenzhen Science and Technology Program (JCYJ20220818101217037, KQTD20200820113106007, JCYJ20240813151223031), and the Shenzhen Natural Science Foundation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNan Yan and Shuzhi Zhao conceived the study. Nan Yan, Shuzhi Zhao, and Yi Guo designed the methodology and analysis plan. Xin Jiang, Chongyuan Lian, and Xue Shi conducted data collection and preprocessing. Nan Yan and Shuzhi Zhao developed the machine learning models and performed statistical analysis. Nan Yan, Hanjun Liu, and Shuzhi Zhao interpreted the results. Nan Yan drafted the initial manuscript. All authors critically reviewed and approved the final version. Nan Yan and Yi Guo had full access to all study data and verified its integrity. All authors had final responsibility for the decision to submit for publication.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eL\u0026ouml;we B\u003cem\u003e, et al.\u003c/em\u003e Persistent physical symptoms: definition, genesis, and management. \u003cem\u003eThe Lancet\u003c/em\u003e \u003cstrong\u003e403\u003c/strong\u003e, 2649-2662 (2024).\u003c/li\u003e\n\u003cli\u003eL\u0026ouml;we B\u003cem\u003e, et al.\u003c/em\u003e Somatic symptom disorder: a scoping review on the empirical evidence of a new diagnosis. \u003cem\u003ePsychological Medicine\u003c/em\u003e \u003cstrong\u003e52\u003c/strong\u003e, 632-648 (2022).\u003c/li\u003e\n\u003cli\u003eTozzi L\u003cem\u003e, et al.\u003c/em\u003e Personalized brain circuit scores identify clinically distinct biotypes in depression and anxiety. \u003cem\u003eNature medicine\u003c/em\u003e \u003cstrong\u003e30\u003c/strong\u003e, 2076-2087 (2024).\u003c/li\u003e\n\u003cli\u003eKroenke K, Spitzer RL, Williams JB. The PHQ-15: validity of a new measure for evaluating the severity of somatic symptoms. \u003cem\u003ePsychosomatic medicine\u003c/em\u003e \u003cstrong\u003e64\u003c/strong\u003e, 258-266 (2002).\u003c/li\u003e\n\u003cli\u003eToussaint A\u003cem\u003e, et al.\u003c/em\u003e Development and Validation of the Somatic Symptom Disorder-B Criteria Scale (SSD-12). \u003cem\u003ePsychosom Med\u003c/em\u003e \u003cstrong\u003e78\u003c/strong\u003e, 5-12 (2016).\u003c/li\u003e\n\u003cli\u003eZeng S\u003cem\u003e, et al.\u003c/em\u003e Neuro-11: a new questionnaire for the assessment of somatic symptom disorder in general hospitals. \u003cem\u003eGeneral Psychiatry\u003c/em\u003e \u003cstrong\u003e36\u003c/strong\u003e, (2023).\u003c/li\u003e\n\u003cli\u003eBogaerts K\u003cem\u003e, et al.\u003c/em\u003e Brain mediators of negative affect-induced physical symptom reporting in patients with functional somatic syndromes. \u003cem\u003eTranslational psychiatry\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 285 (2023).\u003c/li\u003e\n\u003cli\u003eKas MJ\u003cem\u003e, et al.\u003c/em\u003e Precision psychiatry roadmap: towards a biology-informed framework for mental disorders. \u003cem\u003eMolecular psychiatry\u003c/em\u003e, 1-10 (2025).\u003c/li\u003e\n\u003cli\u003eHallberg H, Maroti D, Lumley MA, Johansson R. Internet-delivered emotional awareness and expression therapy for somatic symptom disorder: one year follow-up. \u003cem\u003eFrontiers in Psychiatry\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, 1505318 (2025).\u003c/li\u003e\n\u003cli\u003eZhang J\u003cem\u003e, et al.\u003c/em\u003e Cortical and subcortical mapping of the human allostatic-interoceptive system using 7 Tesla fMRI. \u003cem\u003eNat Neurosci\u003c/em\u003e, (2025).\u003c/li\u003e\n\u003cli\u003eSeymour B, Crook RJ, Chen ZS. Post-injury pain and behaviour: a control theory perspective. \u003cem\u003eNature reviews Neuroscience\u003c/em\u003e \u003cstrong\u003e24\u003c/strong\u003e, 378-392 (2023).\u003c/li\u003e\n\u003cli\u003eZhang Y, Chen ZS. Harnessing electroencephalography connectomes for cognitive and clinical neuroscience. \u003cem\u003eNat Biomed Eng\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, 1186-1201 (2025).\u003c/li\u003e\n\u003cli\u003eGrabowski L B, Chelstowski M, Hiszpanska M, Laszewska K, Lewandowska M, Milner R. Abnormalities in the absolute power of Delta and Alpha rhythms in the frontal lobe of patients suffering from psychosomatic disorders. \u003cem\u003ePsychiatria polska\u003c/em\u003e, 1-12 (2024).\u003c/li\u003e\n\u003cli\u003eHong JK, Park HY, Yoon IY, Jang YE. Longitudinal qEEG changes correlate with clinical outcomes in patients with somatic symptom disorder. \u003cem\u003eJ Psychosom Res\u003c/em\u003e \u003cstrong\u003e151\u003c/strong\u003e, 110637 (2021).\u003c/li\u003e\n\u003cli\u003eNovelli L, Razi A. A mathematical perspective on edge-centric brain functional connectivity. \u003cem\u003eNat Commun\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 2693 (2022).\u003c/li\u003e\n\u003cli\u003eLiu B\u003cem\u003e, et al.\u003c/em\u003e Self-supervised learning reveals clinically relevant histomorphological patterns for therapeutic strategies in colon cancer. \u003cem\u003eNat Commun\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e, 2328 (2025).\u003c/li\u003e\n\u003cli\u003eRafiei MH, Gauthier LV, Adeli H, Takabi D. Self-Supervised Learning for Electroencephalography. \u003cem\u003eIEEE Trans Neural Netw Learn Syst\u003c/em\u003e \u003cstrong\u003e35\u003c/strong\u003e, 1457-1471 (2024).\u003c/li\u003e\n\u003cli\u003eBrucar LR, Feczko E, Fair DA, Zilverstand A. Current approaches in computational psychiatry for the data-driven identification of brain-based subtypes. \u003cem\u003eBiological psychiatry\u003c/em\u003e \u003cstrong\u003e93\u003c/strong\u003e, 704-716 (2023).\u003c/li\u003e\n\u003cli\u003eWang X, Xu Y. An improved index for clustering validation based on Silhouette index and Calinski-Harabasz index. In: \u003cem\u003eIOP Conference Series: Materials Science and Engineering\u003c/em\u003e). IOP Publishing (2019).\u003c/li\u003e\n\u003cli\u003eLai M, Demuru M, Hillebrand A, Fraschini M. A comparison between scalp- and source-reconstructed EEG networks. \u003cem\u003eSci Rep\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e, 12269 (2018).\u003c/li\u003e\n\u003cli\u003eKimes PK, Liu Y, Neil Hayes D, Marron JS. Statistical significance for hierarchical clustering. \u003cem\u003eBiometrics\u003c/em\u003e \u003cstrong\u003e73\u003c/strong\u003e, 811-821 (2017).\u003c/li\u003e\n\u003cli\u003eHelmer M\u003cem\u003e, et al.\u003c/em\u003e On the stability of canonical correlation analysis and partial least squares with application to brain-behavior associations. \u003cem\u003eCommunications biology\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, 217 (2024).\u003c/li\u003e\n\u003cli\u003eYizhar O, Fenno LE, Davidson TJ, Mogri M, Deisseroth K. Optogenetics in neural systems. \u003cem\u003eNeuron\u003c/em\u003e \u003cstrong\u003e71\u003c/strong\u003e, 9-34 (2011).\u003c/li\u003e\n\u003cli\u003eCanolty RT, Knight RT. The functional role of cross-frequency coupling. \u003cem\u003eTrends in cognitive sciences\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 506-515 (2010).\u003c/li\u003e\n\u003cli\u003ePetzschner FH, Garfinkel SN, Paulus MP, Koch C, Khalsa SS. Computational models of interoception and body regulation. \u003cem\u003eTrends in neurosciences\u003c/em\u003e \u003cstrong\u003e44\u003c/strong\u003e, 63-76 (2021).\u003c/li\u003e\n\u003cli\u003eMulders D, Seymour B, Mouraux A, Mancini F. Confidence of probabilistic predictions modulates the cortical response to pain. \u003cem\u003eProceedings of the National Academy of Sciences\u003c/em\u003e \u003cstrong\u003e120\u003c/strong\u003e, e2212252120 (2023).\u003c/li\u003e\n\u003cli\u003e\u0026Ouml;zt\u0026uuml;rk S, Devecioglu I, G\u0026uuml;\u0026ccedil;l\u0026uuml; B. Bayesian prediction of psychophysical detection responses from spike activity in the rat sensorimotor cortex. \u003cem\u003eJournal of computational neuroscience\u003c/em\u003e \u003cstrong\u003e51\u003c/strong\u003e, 207-222 (2023).\u003c/li\u003e\n\u003cli\u003eUddin LQ, Kinnison J, Pessoa L, Anderson ML. Beyond the tripartite cognition-emotion-interoception model of the human insular cortex. \u003cem\u003eJournal of cognitive neuroscience\u003c/em\u003e \u003cstrong\u003e26\u003c/strong\u003e, 16-27 (2014).\u003c/li\u003e\n\u003cli\u003eChen WG\u003cem\u003e, et al.\u003c/em\u003e The Emerging Science of Interoception: Sensing, Integrating, Interpreting, and Regulating Signals within the Self. \u003cem\u003eTrends Neurosci\u003c/em\u003e \u003cstrong\u003e44\u003c/strong\u003e, 3-16 (2021).\u003c/li\u003e\n\u003cli\u003eUddin LQ. Salience processing and insular cortical function and dysfunction. \u003cem\u003eNature reviews neuroscience\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e, 55-61 (2015).\u003c/li\u003e\n\u003cli\u003eVan den Bergh O, Witth\u0026ouml;ft M, Petersen S, Brown RJ. Symptoms and the body: taking the inferential leap. \u003cem\u003eNeuroscience \u0026amp; Biobehavioral Reviews\u003c/em\u003e \u003cstrong\u003e74\u003c/strong\u003e, 185-203 (2017).\u003c/li\u003e\n\u003cli\u003eZhang R, Deng H, Xiao X. The Insular Cortex: An Interface Between Sensation, Emotion and Cognition. \u003cem\u003eNeuroscience Bulletin\u003c/em\u003e, 1-11 (2024).\u003c/li\u003e\n\u003cli\u003eAkiki TJ, Jubeir J, Bertrand C, Tozzi L, Williams LM. Neural circuit basis of pathological anxiety. \u003cem\u003eNature reviews Neuroscience\u003c/em\u003e \u003cstrong\u003e26\u003c/strong\u003e, 5-22 (2025).\u003c/li\u003e\n\u003cli\u003eSimmons WK, DeVille DC. Interoceptive contributions to healthy eating and obesity. \u003cem\u003eCurrent opinion in psychology\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, 106-112 (2017).\u003c/li\u003e\n\u003cli\u003eWiebking C, Bauer A, de Greck M, Duncan NW, Tempelmann C, Northoff G. Abnormal body perception and neural activity in the insula in depression: an fMRI study of the depressed \u0026quot;material me\u0026quot;. \u003cem\u003eThe world journal of biological psychiatry : the official journal of the World Federation of Societies of Biological Psychiatry\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 538-549 (2010).\u003c/li\u003e\n\u003cli\u003eArminjon M, Preissmann D, Chmetz F, Duraku A, Ansermet F, Magistretti PJ. Embodied memory: unconscious smiling modulates emotional evaluation of episodic memories. \u003cem\u003eFrontiers in psychology\u003c/em\u003e \u003cstrong\u003e6\u003c/strong\u003e, 650 (2015).\u003c/li\u003e\n\u003cli\u003eBai J, Kong S, Gomes C. Disentangled variational autoencoder based multi-label classification with covariance-aware multivariate probit model. In: \u003cem\u003eProceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence\u003c/em\u003e) (2021).\u003c/li\u003e\n\u003cli\u003eBarroso J, Branco P, Apkarian AV. The causal role of brain circuits in osteoarthritis pain. \u003cem\u003eNature Reviews Rheumatology\u003c/em\u003e, 1-14 (2025).\u003c/li\u003e\n\u003cli\u003eChang LJ, Yarkoni T, Khaw MW, Sanfey AG. Decoding the role of the insula in human cognition: functional parcellation and large-scale reverse inference. \u003cem\u003eCerebral cortex (New York, NY : 1991)\u003c/em\u003e \u003cstrong\u003e23\u003c/strong\u003e, 739-749 (2013).\u003c/li\u003e\n\u003cli\u003eLi C\u003cem\u003e, et al.\u003c/em\u003e Classification of schizophrenia spectrum disorder using machine learning and functional connectivity: reconsidering the clinical application. \u003cem\u003eBMC Psychiatry\u003c/em\u003e \u003cstrong\u003e25\u003c/strong\u003e, 372 (2025).\u003c/li\u003e\n\u003cli\u003eLiang Z, Ye W, Liu Q, Zhang L, Huang G, Zhou Y. NSSI-Net: A Multi-Concept GAN for Non-Suicidal Self-Injury Detection Using High-Dimensional EEG in a Semi-Supervised Framework. (2024).\u003c/li\u003e\n\u003cli\u003eKrueger RF\u003cem\u003e, et al.\u003c/em\u003e Progress in achieving quantitative classification of psychopathology. \u003cem\u003eWorld Psychiatry\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, 282-293 (2018).\u003c/li\u003e\n\u003cli\u003eThut G\u003cem\u003e, et al.\u003c/em\u003e Guiding transcranial brain stimulation by EEG/MEG to interact with ongoing brain activity and associated functions: A position paper. \u003cem\u003eClinical neurophysiology : official journal of the International Federation of Clinical Neurophysiology\u003c/em\u003e \u003cstrong\u003e128\u003c/strong\u003e, 843-857 (2017).\u003c/li\u003e\n\u003cli\u003eZhao A, Fattahi D, Hu X. Physics-informed neural networks for physiological signal processing and modeling: a narrative review. \u003cem\u003ePhysiological measurement\u003c/em\u003e \u003cstrong\u003e46\u003c/strong\u003e, (2025).\u003c/li\u003e\n\u003cli\u003eAouedi O, Sacco A, Piamrat K, Marchetto G. Handling Privacy-Sensitive Medical Data With Federated Learning: Challenges and Future Directions. \u003cem\u003eIEEE journal of biomedical and health informatics\u003c/em\u003e \u003cstrong\u003e27\u003c/strong\u003e, 790-803 (2023).\u003c/li\u003e\n\u003cli\u003eKop WJ, Toussaint A, Mols F, L\u0026ouml;we B. Somatic symptom disorder in the general population: Associations with medical status and health care utilization using the SSD-12. \u003cem\u003eGen Hosp Psychiatry\u003c/em\u003e \u003cstrong\u003e56\u003c/strong\u003e, 36-41 (2019).\u003c/li\u003e\n\u003cli\u003eBagby RM, Ryder AG, Schuller DR, Marshall MB. The Hamilton Depression Rating Scale: has the gold standard become a lead weight? \u003cem\u003eAmerican Journal of Psychiatry\u003c/em\u003e \u003cstrong\u003e161\u003c/strong\u003e, 2163-2177 (2004).\u003c/li\u003e\n\u003cli\u003eLeentjens AF, Dujardin K, Marsh L, Richard IH, Starkstein SE, Martinez-Martin P. Anxiety rating scales in Parkinson\u0026apos;s disease: a validation study of the Hamilton anxiety rating scale, the Beck anxiety inventory, and the hospital anxiety and depression scale. \u003cem\u003eMovement Disorders\u003c/em\u003e \u003cstrong\u003e26\u003c/strong\u003e, 407-415 (2011).\u003c/li\u003e\n\u003cli\u003ePetropoulakos K\u003cem\u003e, et al.\u003c/em\u003e Validity and reliability of the Greek version of Pittsburgh Sleep Quality Index in chronic non-specific low back pain patients. In: \u003cem\u003eHealthcare\u003c/em\u003e). MDPI (2024).\u003c/li\u003e\n\u003cli\u003eDelorme A, Makeig S. EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysis. \u003cem\u003eJournal of neuroscience methods\u003c/em\u003e \u003cstrong\u003e134\u003c/strong\u003e, 9-21 (2004).\u003c/li\u003e\n\u003cli\u003eGramfort A\u003cem\u003e, et al.\u003c/em\u003e MEG and EEG data analysis with MNE-Python. \u003cem\u003eFrontiers in Neuroinformatics\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, 267 (2013).\u003c/li\u003e\n\u003cli\u003eNoack MM\u003cem\u003e, et al.\u003c/em\u003e Gaussian processes for autonomous data acquisition at large-scale synchrotron and neutron facilities. \u003cem\u003eNature Reviews Physics\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, 685-697 (2021).\u003c/li\u003e\n\u003cli\u003eBai J, Kong S, Gomes CP. Gaussian mixture variational autoencoder with contrastive learning for multi-label classification. In: \u003cem\u003einternational conference on machine learning\u003c/em\u003e). PMLR (2022).\u003c/li\u003e\n\u003cli\u003eFranco-Pereira AM, Nakas CT, Reiser B, Pardo MC. Inference on the overlap coefficient: the binormal approach and alternatives. \u003cem\u003eStatistical methods in medical research\u003c/em\u003e \u003cstrong\u003e30\u003c/strong\u003e, 2672-2684 (2021).\u003c/li\u003e\n\u003cli\u003eLiu Y, Hayes DN, Nobel A, Marron JS. Statistical significance of clustering for high-dimension, low\u0026ndash;sample size data. \u003cem\u003eJournal of the American Statistical Association\u003c/em\u003e \u003cstrong\u003e103\u003c/strong\u003e, 1281-1293 (2008).\u003c/li\u003e\n\u003cli\u003eZamani Esfahlani F, Byrge L, Tanner J, Sporns O, Kennedy DP, Betzel RF. Edge-centric analysis of time-varying functional brain networks with applications in autism spectrum disorder. \u003cem\u003eNeuroimage\u003c/em\u003e \u003cstrong\u003e263\u003c/strong\u003e, 119591 (2022).\u003c/li\u003e\n\u003cli\u003eYeo BT\u003cem\u003e, et al.\u003c/em\u003e The organization of the human cerebral cortex estimated by intrinsic functional connectivity. \u003cem\u003eJournal of neurophysiology\u003c/em\u003e, (2011).\u003c/li\u003e\n\u003cli\u003eArzate-Mena JD\u003cem\u003e, et al.\u003c/em\u003e Stationary EEG pattern relates to large-scale resting state networks\u0026ndash;An EEG-fMRI study connecting brain networks across time-scales. \u003cem\u003eNeuroImage\u003c/em\u003e \u003cstrong\u003e246\u003c/strong\u003e, 118763 (2022).\u003c/li\u003e\n\u003cli\u003ePu Y\u003cem\u003e, et al.\u003c/em\u003e Variational autoencoder for deep learning of images, labels and captions. \u003cem\u003eAdvances in neural information processing systems\u003c/em\u003e \u003cstrong\u003e29\u003c/strong\u003e, (2016).\u003c/li\u003e\n\u003cli\u003eWu M, Goodman N. Multimodal generative models for scalable weakly-supervised learning. \u003cem\u003eAdvances in neural information processing systems\u003c/em\u003e \u003cstrong\u003e31\u003c/strong\u003e, (2018).\u003c/li\u003e\n\u003cli\u003eYi K, Wang Y, Ren K, Li D. Learning topology-agnostic EEG representations with geometry-aware modeling. In: \u003cem\u003eProceedings of the 37th International Conference on Neural Information Processing Systems\u003c/em\u003e). Curran Associates Inc. (2023).\u003c/li\u003e\n\u003cli\u003eJiang WB, Zhao LM, Lu BL. Large Brain Model for Learning Generic Representations with Tremendous EEG Data in BCI. (2024).\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTable 1. Performance Comparison of Representative Models under Model-Driven and Data-Driven Strategies\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003eModels\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 409px;\"\u003e\n \u003cp\u003eEvaluation Criterion\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003eACC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003eAUC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003eSensitivity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e1-Specificity\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003eVAE \u003csup\u003e59\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e0.77\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e0.79\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e0.76\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e0.78\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003eMPVAE \u003csup\u003e37\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e0.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e0.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e0.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e0.80\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003ePOE-VAE \u003csup\u003e60\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e0.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e0.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e0.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e0.82\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003eVAE-GM \u003csup\u003e59\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e0.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e0.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e0.81\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e0.82\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003eMPVAE-GM \u003csup\u003e37\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e0.81\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e0.83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e0.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e0.84\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003ePOE-VAE-GM \u003csup\u003e60\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e0.83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e0.85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e0.84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e0.86\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003eNSSINet \u003csup\u003e41\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e0.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e0.83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e0.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e0.81\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003eMMM \u003csup\u003e61\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e0.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e0.84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e0.81\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e0.83\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003eLaBraM \u003csup\u003e62\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e0.83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e0.84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e0.85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e0.87\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003eCVAE-GM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.85\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.87\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.86\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 106px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.88\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eGM: Gaussian Mixture\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-7971302/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7971302/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"Substantial clinical heterogeneity in Somatic Symptom Disorder (SSD) limits treatment efficacy. Here， this study propose a data-driven framework using neurophysiology information to identify distinct patient subtypes. A contrastive variational autoencoder with Gaussian mixture modeling (CVAE-GM) was developed using resting-state EEG connectomics from a discovery cohort of 1,419 patients with SSD. The derived subtypes were clinically correlated with symptom dimensions and validated for reproducibility in an independent external cohort (n=530). We identified three robust and reproducible subtypes, characterized by dominant connectivity in somatomotor, central executive, or limbic networks with insula area. Each neurophysiological subtype was significantly associated with distinct clinical profiles (somatic symptoms, adverse cognitive, and negative emotions). The subtyping model demonstrated high reproducibility in the validation cohort (accuracy = 0.85; AUC = 0.87). This EEG-based framework provides a validated, neurobiological basis for stratifying SSD patients. It transcends symptom-based nosology, enabling a precision medicine approach for developing targeted interventions.","manuscriptTitle":"Multi-Frequency EEG Connectomics Uncovers Insula-Network Subtypes in Somatic Symptom Disorder","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-11-04 07:47:26","doi":"10.21203/rs.3.rs-7971302/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"cd3798aa-8ea6-40bf-8122-3f9ed2bc2c44","owner":[],"postedDate":"November 4th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":57182249,"name":"Health sciences/Biomarkers/Diagnostic markers"},{"id":57182250,"name":"Biological sciences/Computational biology and bioinformatics/Machine learning"},{"id":57182251,"name":"Health sciences/Diseases/Psychiatric disorders"},{"id":57182252,"name":"Biological sciences/Biological techniques/Electrophysiology/Electroencephalography \u0026#x2013; EEG"},{"id":57182253,"name":"Biological sciences/Neuroscience/Somatosensory system/Insula"}],"tags":[],"updatedAt":"2025-11-10T21:15:36+00:00","versionOfRecord":[],"versionCreatedAt":"2025-11-04 07:47:26","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7971302","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7971302","identity":"rs-7971302","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}