EEG-based classification has been extensively employed in the literature to differentiate psychiatric disorders, but challenges in replicability, generalizability, and improving the sensitivity and specificity of classification methods persist in identifying objective neural biomarkers. In this study, we aimed to address these challenges by conducting a two-layer, four-way classification of participants, including healthy individuals and patients with two subtypes of depression (NPMD and PMD) and SCZ.
Our findings revealed that the identification of four templates of brain states as the fundamental building block of single-subject EEG time courses was robustly confirmed, demonstrating within-diagnosis and across-diagnosis reproducibility. Furthermore, we demonstrated the ability of DFC features derived from subject-specific brain state sequences to differentiate between patients with depression subtypes and SCZ, exhibiting high classification performance. Interestingly, our results also indicated that the DFC approach may capture important information not captured by the SFC approach, yielding higher overall accuracy rates. However, the combination of SFC and DFC features in our classification approach only yielded a slight improvement in overall accuracy, suggesting that SFC features may contribute little significant information. Moreover, utilizing a single classifier and the Boruta algorithm for four-way classification resulted in an accuracy of 63.5%, slightly lower than the two-layer classification approach (75.7%) proposed in our study. To the best of our knowledge, our study is the first to distinguish NPMD, PMD, and SCZ using resting-state EEG DFC features and a two-layer random forest classification algorithm, achieving satisfactory sensitivity and accuracy. These findings contribute objective EEG biological markers for the early identification of common psychiatric disorders, such as depression, schizophrenia, and psychotic depression, which holds great significance for clinical diagnosis and forensic psychiatry.
In the field of psychiatric research, there have been efforts to explore the relationship between psychiatric symptoms, such as depressive symptoms (e.g., low mood, lack of motivation) or thinking styles (e.g., worrying), and reduce the misdiagnosis of schizophrenia, depression, and psychotic depression based solely on clinical presentation and diagnostic criteria [41, 42]. However, inconsistencies in symptom thresholds and subtypes across different guidelines and diagnostic criteria have contributed to diagnosis discrepancies and hindered the comparison of research findings, resulting in delayed treatment, poor prognosis, and negative societal consequences [43]. Therefore, the high classification accuracy achieved for the three diseases (NPMD, PMD, and SCZ) in our study provides an important reference for early identification in clinical practice, emphasizing the significance of EEG technology in psychiatric disease diagnosis.
Furthermore, the study of dynamic properties of functional connectivity has gained increasing attention in understanding whole brain activities [30] and their relationship to cognition and behavior [37, 44]. However, most DFC studies have primarily focused on group-level analyses [13, 23] and have not adequately demonstrated reliability at the individual level. Therefore, further research is needed to assess the reliability of spatiotemporal brain network patterns and understand their reproducibility across individuals and over time. In our study, we confirmed the stability of estimated brain states within and across different diagnostic groups, which is a crucial step in linking transient brain states to individual differences in brain diseases. Our DFC approach provides a powerful analytical methodology for identifying network-dynamic commonalities and constructing temporal representations of brain state dynamics at the individual level. Overall, our findings demonstrate the reliability of within-diagnosis and across-diagnosis brain states and capture individual differences in the temporal domain of functional brain activity.
In the present study, we made an intriguing finding regarding significant differences in two brain states across different diagnostic groups. State 2 was found to involve regions of the default mode network (DMN), including electrodes placed over midline areas of the frontal, central, and parietal lobes. These locations are associated with the medial prefrontal cortex (mPFC), posterior cingulate cortex (PCC), and precuneus. On the other hand, State 3 appeared to involve regions of the frontoparietal network (FPN), including electrodes FP1, F3, P3, P4, and P5. These locations correspond to the dorsolateral prefrontal cortex (DLPFC) and posterior parietal cortex (PPC). Furthermore, we observed that the proportion of State 2 increased in the NPMD, PMD, and SCZ groups, while the proportion of State 3 decreased (Fig. 4). Previous meta-analytic studies [45] have characterized MDD by hyperconnectivity within the DMN, a network associated with self-referential thought [46], and hypoconnectivity within the FPN, a network involved in attention and emotion regulation [45, 47]. Based on these findings, we speculate that individuals with MDD would spend more time in State 2, which is associated with enhanced positive connectivity in the DMN, and less time in State 3, which is associated with weaker positive connectivity in the FPN. We also propose that similar patterns may be observed in PMD and SCZ, suggesting comparable outcomes across these disorders. Our results align with previous meta-analytic studies [45] indicating abnormal communication among functional networks in psychiatric disorders. Specifically, deficits in cognitive control, commonly observed in depression, may be attributed to abnormal communication within the FPN, potentially contributing to symptoms such as difficulty concentrating or regulating emotions [45]. Additionally, our study expands upon prior research by demonstrating alterations in the dynamic spatiotemporal profiles of brain networks in depression and schizophrenia, in addition to static network connectivity patterns. However, to establish a stronger link between our findings and existing literature, and to hypothesize how altered dynamic configurations relate to MDD and SCZ symptoms, it is crucial to collect symptomatic profiles from MDD and SCZ patients. By systematically exploring the dynamics of State 2 and State 3 in relation to the severity and presence of different symptoms, we can enhance our understanding of these disorders and potentially inform future diagnostic and treatment strategies.
Regarding the spectral findings, we identified three frequency bands (delta, theta, and gamma) that showed significant differences among the four groups (HC, NPMD, PMD, and SCZ) in the DFC features. Synchronous oscillations in these frequency bands have been associated with cognitive functions such as selective attention, working memory, emotional regulation, and consciousness [48]. Previous studies have implicated abnormalities in gamma band synchrony in the prefrontal cortex in MDD and SCZ [49, 50], while the exact relationship between theta band rhythms and MDD or SCZ is still an area of ongoing research. Our study, with its large sample size and consideration of multiple subtypes of psychiatric disorders, provides valuable insights and helps reconcile inconsistencies in the delta rhythms reported in prior research. The robustness and statistical power derived from our approach enhance the reliability of our findings, providing a more comprehensive understanding of the synchronous oscillations of EEG in psychiatric disorders. However, the observed discrepancies in the literature may be attributed to variations in methodologies, sample characteristics, and measurement tools across different studies. Therefore, ongoing exploration and result replication are crucial for establishing a clearer comprehension of the complexities inherent in psychiatric disorders. Our study serves as a significant reference, shedding light on the need for further research to expand upon and refine our understanding of psychiatric disorders. By building upon our findings, future investigations can deepen our knowledge and pave the way for advancements in this field.
Limitations
This study had the following limitations. First, the retrospective cross-sectional study design and the small number of samples in the PMD group may lead to deviations in the results. Second, some of the depression and schizophrenia patients in this study had received drug treatment, which may induce unexpected variation in the heterogeneity of the patient groups. Moreover, the data used in this study was collected from the same institution, and the identified aberrant dynamic FC states were statistically linked to our data. Though we observed a convergence of state similarity across the groups, further methodological validation with more diverse and multi-center data is necessary.