Continuous variables were expressed as mean ± standard deviation (SD), while categorical variables were summarized as counts and percentages. The normality of continuous variables was assessed using the Shapiro–Wilk test, and homogeneity of variance was determined via the Levene test. Group comparisons for categorical variables were performed using the chi-square test. For continuous variables, independent-sample t tests were used when the data met assumptions of normality and homogeneity of variance; otherwise, the nonparametric Mann–Whitney U test was applied. For comparisons involving 3 or more groups, either one-way analysis of variance or the Kruskal–Wallis H test was employed based on data distribution, followed by Tukey’s or Dunn’s post hoc tests when necessary. When multiple comparisons were performed, P values were adjusted using the Benjamini–Hochberg method to control the false discovery rate. All statistical analyses were conducted using R or SPSS software, with P value < 0.05 considered statistically significant.

Over the course of 6 years, we established a dataset comprising 1,266 participants, including 893 patients with psychiatric disorders: 509 with MDD (of whom 230 were treatment-naive, first-episode patients), 63 with BD, and 321 with acute exacerbations of SZ. The mean disease duration among patients was 53 months. Additionally, 373 HC were recruited. Detailed demographic and clinical characteristics are summarized in Table 2. Additionally, we conducted structured assessments of participants’ lifestyle habits and health status at both baseline and follow-up stages. These assessments included information on medication adherence, dietary patterns, gastrointestinal symptoms, antibiotic use history, daily activity levels, and exercise frequency. The incorporation of these scales facilitates effective control for lifestyle confounders in subsequent analyses. The relevant follow-up questionnaires and scale contents are fully documented in the Supplementary Materials.

BIGHI is grounded in real-world clinical cohorts and is designed to support the systematic expansion of sample size and accumulation of multimodal data within an established framework for cross-system and cross-diagnostic joint modeling. Core clinical assessments and symptom rating scales are uniformly collected across all participants, with EEG, MRI, gut microbiome profiles, and peripheral biomarkers forming the primary multimodal data layers. Longitudinal follow-up is conducted around key clinical time windows using a structured yet flexible design compatible with real-world clinical practice. Further details regarding sample size planning, multimodal overlap, and follow-up design are provided in the Supplementary Materials.

From an electrophysiological perspective, EEG enables the characterization of dynamic neural activity patterns in individuals with psychiatric disorders, facilitating the identification of aberrant neural activity features and potential electrophysiological correlates. Through EEG microstate analysis, we evaluated the therapeutic effects of repetitive transcranial magnetic stimulation (rTMS) on positive symptoms in chronic SZ patients [30]. Over a 4-week treatment period, patients received active or sham 10 Hz stimulation targeting the left dorsolateral prefrontal cortex (DLPFC) for 5 sessions per week. Active rTMS resulted in a significant reduction in the duration of microstate C and a marked increase in microstate D, indicating that 10 Hz stimulation can ameliorate positive symptoms in chronic SZ and that EEG microstates may serve as sensitive indicators of symptom improvement (Fig. 3A). Furthermore, we calculated features such as nonlinear dynamics across different frequency bands of EEG and brain functional network properties, analyzed EEG signals of SZ patients under cognitive load, and revealed the association between frontal lobe functional impairment and cognitive deficits in SZ. We constructed an automatic classification method based on these EEG features, achieving an accuracy rate of 76.77%, a sensitivity of 72.09%, and a specificity of 80.36% in classifying SZ patients (Fig. 3B).

In studies of MDD, integrating EEG spectral features with machine learning has identified an abnormal pattern of cortical activity that is thought to indirectly reflect dysregulated cortical excitability. MDD patients showed markedly reduced relative α-band power across the parieto-occipital region, attenuated scalp functional connectivity in the α band, and enhanced γ-band connectivity. Notably, reduced α-band power in the left parietal region emerged as a key electrophysiological marker of MDD [31] (Fig. 3C). Recently, by comparing EEG data from patients with MDD and HC, we observed that in the MDD group, there was a significant reduction in alpha wave power in the parieto-occipital region and beta wave power in the right parieto-occipital area. Moreover, the nonperiodic intercept in the parieto-occipital region of the MDD group was significantly lower and positively correlated with the HAMD scores (Fig. 3D), suggesting that this nonperiodic component may serve as a potential biomarker for depression [32]. Additionally, we observed systematic changes in EEG microstate features across different severity levels of MDD symptoms. MDD patients generally exhibited reduced microstate A and increased microstate C, while those with severe MDD also demonstrated prolonged microstate D, suggesting compensatory overactivation of the executive control network. As symptoms worsened, microstate transition flexibility further declined, and microstate D emerged as a potential neural indicator of depression severity [33]. When comparing MDD patients with and without suicidal ideation (SI) to HC, a significant reduction in microstate D was observed, manifested as decreased coverage and fewer transitions from A to D. These alterations accompany slowed processing speed, suggesting that impaired cognitive control functions may represent a key neural mechanism underlying SI [34]. Collectively, these findings highlight EEG’s potential for identifying brain abnormalities and aiding psychiatric diagnosis.

Through the integration of structural and functional MRI, we systematically mapped brain network alterations across diagnostic groups, aiming to pinpoint key regions and connectivity features linked to symptoms, cognition, and illness stage. In terms of SZ, we constructed brain functional networks in patients following antipsychotic treatment and found that changes in the degree centrality of the left anterior inferior parietal lobule were negatively correlated with changes in positive scores (Fig. 4A). Furthermore, the post-treatment degree centrality of this region was significantly negatively associated with performance on the Symbol Digit Modalities Test [35]. Subsequently, we constructed structural–functional brain connectivity maps for first-episode drug-naive SZ, chronic SZ, and HC, and calculated the structural–functional (SC–FC) coupling strength. The results showed that, at the global connectivity level, chronic SZ patients had significantly higher SC–FC coupling strength than first-episode patients. At the node-strength level, first-episode SZ patients exhibited markedly lower coupling strength than HC, and this reduction was positively correlated with PANSS negative symptom scores [36] (Fig. 4B). In addition, by calculating dynamic effective connectivity, we also found that the dynamic connectivity patterns between the cerebellar dentate nucleus and the cerebral cortex in SZ patients were significantly different from those in HC, exhibiting a marked reduction in transition frequency (Fig. 4C), which was associated with neurocognitive functions [37]. As part of our work on intelligent diagnosis, we computed key MRI-based features, including amplitude of low-frequency fluctuations (ALFF), regional homogeneity (ReHo), and degree centrality, and further analyzed large-scale functional connectivity patterns. Using these multimodal features, we constructed a GCN model to classify SZ and HC. Compared with traditional machine learning and deep learning approaches that rely solely on regional MRI features, integrating functional MRI with connectomics markedly improved classification performance. The model achieved an accuracy of 92.47%, with areas under the receiver operating curve (AUC), sensitivity, specificity, precision, and F1 score of 95.36%, 88.70%, 95.06%, 92.87%, and 90.45%, respectively [38]. Additionally, we classified SZ patients and HC using 3 brain atlases and 5 classifiers, validating the classification performance with 2 cross-validation methods. The results indicated that the whole-brain atlas with 268 ROIs outperformed the other 2 brain atlases [39]. We also classified SZ patients and HC using an SVM model based on gray matter volume (GMV) and white matter features, achieving an accuracy of 88.4%. Notably, reductions in GMV and alterations in white matter among SZ patients were primarily observed in regions implicated in emotion, memory, and visual processing [40]. Furthermore, we developed an integrated deep learning model that combines 3D and 2D convolutional neural network (CNN) architectures to classify SZ patients using multimodal MRI features. In this framework, 3D images of T1, ReHo, and ALFF were used as inputs to the 3D CNN, whereas SC and FC matrices were used as inputs to the 2D CNN. SE blocks were incorporated, and the softmax layer was replaced with an SVM classifier. The resulting integrated model achieved an accuracy of 89.86%, demonstrating strong potential for improving automated SZ diagnosis [41] (Fig. 4D).

We also conducted studies on BD and MDD. Using resting-state functional connectivity analysis, we found that patients with BD who exhibited suicidal ideation showed significantly reduced voxel-wise consistency in the left fusiform gyrus, left lingual gyrus, right middle temporal gyrus, and left postcentral gyrus compared to those without suicidal ideation. Additionally, in patients with suicidal ideation (Fig. 4E), voxel-wise consistency in the left middle occipital gyrus was negatively correlated with working memory performance [42]. In addition, we examined the relationship among plasma homocysteine (Hcy) levels, brain function, and cognition in BD patients with suicidal ideation (BDSI) and without suicidal ideation (BDNSI). Hcy levels were significantly lower in the BDSI group and positively correlated with fALFF signals in the left posterior cingulate gyrus (Fig. 4F), which further mediated the negative association between Hcy and visual/verbal learning performance [43]. In adolescent MDD patients, we observed a significant reduction in FC between the left posterior putamen and the right dorsomedial prefrontal cortex compared with HC. Further moderation analysis showed that this aberrant FC modulated the relationship between childhood maltreatment and the severity of adolescent-onset MDD symptoms. Interaction analysis additionally revealed that when FC was lower, childhood maltreatment exhibited a stronger positive association with MDD symptom severity [44] (Fig. 4G). In addition, we constructed an edge-centric functional network and compared node entropy between MDD patients with and without suicidal ideation. Patients with suicidal ideation showed reduced subnet entropy in the dorsal attention network and increased entropy in the default mode network. Node-entropy features effectively distinguished the 2 groups, and the observed alterations were associated with the expression of genes involved in cell signaling and immune-inflammatory processes (Fig. 4H). These alterations were associated with visual learning performance and the expression of genes related to immune and inflammatory responses [45].

At the microbiological level, we comprehensively evaluated gut microbiota alterations in major psychiatric disorders to uncover structured associations among microbial profiles, metabolic processes, and clinical symptoms. We found significant alterations in the gut microbiota of patients with SZ [46], including alterations in β-diversity, increases in specific taxa such as Collinsella and Lactobacillus, and disruptions in multiple metabolic pathways such as pantothenate and CoA biosynthesis. Among these microbial features, the abundance of Succinivibrio was positively associated with symptom severity, whereas Corynebacterium showed a negative association (Fig. 5A). We also collected cognitive function, blood biochemical indicators, and gut microbiome data from patients with SZ, focusing on the relationships between oxidative stress markers such as superoxide dismutase (SOD) and Hcy with disease risk and cognitive decline. Machine learning models were further applied to evaluate the potential of SOD as a biomarker. Our findings showed that peripheral SOD levels were generally elevated in SZ patients and accompanied by pronounced gut microbial dysbiosis. The abundances of Collinsella, Desulfovibrio, and other taxa were increased, whereas beneficial genera such as Faecalibacterium, Anaerostipes, Turicibacter, and Ruminococcus were markedly reduced. SOD levels were systematically associated with the abundances of Eubacterium, Collinsella, and Lactobacillus, as well as with cognitive performance, suggesting that oxidative stress may influence brain function through microbiome-related pathways (Fig. 5B) [47]. Another study further demonstrated that Hcy levels were significantly higher in SZ patients compared with HC and were closely associated with poorer verbal learning performance. In addition, Hcy levels showed positive correlations with several gut microbial taxa, including Lactobacillus, Bifidobacterium, Corynebacterium, Mogibacterium, and Bulleidia (Fig. 5C) [48]. In an obesity-specific study of SZ, significant alterations in gut microbiota composition and metabolic profiles were observed. Notably, increased abundance of Collinsella was significantly negatively associated with cognitive function (Fig. 5D), while decreased levels of Clostridium and Butyricicoccus were positively associated with cognitive function [49]. Moreover, studies focusing on nonbacterial gut microbes remain scarce. Metagenomic analysis revealed pronounced dysbiosis across multiple microbial kingdoms in SZ patients, encompassing bacteria, fungi, archaea, and viruses, and identified 21 significantly altered metabolic pathways that showed strong species-function coordination. In addition, the abundance of Streptococcus vestibularis was positively correlated with Hcy levels (Fig. 5E) [50].

To explore cross-system mechanisms of the MGBA, we further integrated multimodal datasets to identify how interactions across systems jointly shape symptomatology and cognitive function. In extending this framework to SZ, we examined the multidimensional associations between the gut microbiome and brain imaging features, uncovering potential interactions underlying disease manifestations (Fig. 6A). At the genus level, we observed markedly reduced abundances of Ruminococcus and Roseburia and an increased abundance of Veillonella in SZ patients. Consistently, MRI analyses revealed decreased GMV and ReHo, together with increased ALFF across multiple brain regions. Gut microbial α-diversity was significantly linked to both measures, and ReHo in the right superior temporal cortex, left cuneus, and right middle temporal gyrus was negatively associated with the abundance of Roseburia [51]. To further explore the connection between the MGBA and psychiatric symptoms, we performed mediation analysis using gut microbiome as the independent variable. Node properties from 7 brain regions, mainly in visual, language, and motor systems (Fig. 6B), significantly mediated the association between Selinomonas abundance and PANSS scores in SZ patients [52]. In terms of data integration, we established, for the first time, an individualized brain–gut–microbiome network (BGMN) by combining multimodal MRI features with gut microbiome profiles. This model achieved high accuracy in distinguishing SZ patients from HC. Among the most discriminative features, Faecalibacterium was associated with MRI-derived features in visual and subcortical regions, while Collinsella correlated with the default mode and subcortical networks (Fig. 6C). Additionally, working memory performance was positively associated with the connectivity between the right inferior parietal lobule and Collinsella, suggesting the potential of BGMN for biomarker discovery in SZ [53]. Leveraging brain–gut network methodology, we further developed a multi-omics graph convolutional network (MO-GCN) and incorporated an attention mechanism to more effectively integrate multi-omics features. The best classification accuracy reached 84.0%. Interpretability analysis showed that the most informative brain features were primarily related to memory and emotional regulation, such as those in the hippocampus and olfactory cortex, while the key microbial features were mainly from Dorea and Ruminococcus. These important features were significantly associated with clinical disease severity and cognitive function [54] (Fig. 6D). Moreover, to address the marked heterogeneity of SZ, we employed a data-driven approach to identify distinct multi-omics subtypes across brain, gut, and brain–gut integrated features. The results revealed that brain and brain–gut subtypes were more strongly associated with symptom profiles, while gut-derived subtypes exhibited closer links to cognitive function (Fig. 6E), highlighting the multilayered heterogeneity of SZ and offering potential pathways for precision diagnosis and treatment [55]. Building upon the multi-omics phenotyping that reveals the multi-layered heterogeneity of SZ, we further extend our analytical framework to the biological age dimension. Recently, we constructed a biological age prediction model using MRI, gut microbiome, and blood-based data. Unlike traditional single-modality biological age estimates, the multimodal approach substantially improved prediction accuracy. Patients with SZ showed a markedly increased biological age gap, which was closely associated with cognitive impairment and disease severity (Fig. 6F), highlighting its potential value in psychiatric assessment [56]. Finally, by integrating gut microbiome profiles, blood-based biomarkers, and EEG network features into a unified multi-omics framework, we substantially improved the diagnostic performance for SZ. The optimized model achieved an accuracy of 91.7% and an AUC of 96.5%. Notably, the most informative features spanned multiple biological domains, including microbial genera (e.g., Lactobacillus and Prevotella), peripheral biomarkers (SOD and immune cell ratios), and EEG-derived measures of nodal efficiency and local network integration within temporal and frontoparietal regions [57] (Fig. 6G). In studies on MDD, recent research has uncovered a critical coupling mechanism between cognitive impairment in first-episode MDD and gut microbiome dysbiosis. Notably, Amycolatopsis sp. Hca4 exhibits significant correlations with cognitive abilities such as processing speed. It influences working memory by modulating functional connectivity between the medial frontal gyrus and parahippocampal gyrus, suggesting that specific microorganisms may serve as pivotal nodes regulating abnormal brain networks and cognitive impairment [58]. These findings highlight the key role of gut–brain interactions in understanding, diagnosing, and treating major psychiatric disorders.