Electroconvulsive therapy shifting brain dynamics toward more adaptive and integrated states in treatment-resistant depression based on a hidden markov model

Abstract

Treatment-resistant depression (TRD) is characterized by persistent symptoms and poor response to conventional therapies, yet its dynamic neural mechanisms and treatment-related changes remain incompletely understood. This study applied a hidden Markov model (HMM) combined with neurotransmitter atlas mapping to investigate whole-brain spatiotemporal dynamics in TRD and their modulation by electroconvulsive therapy (ECT). Resting-state fMRI data were collected from 72 TRD patients (52 completing follow-up) and 63 healthy controls. HMM identified six recurring brain states defined by distinct patterns of functional connectivity and averaged functional activity. At baseline, TRD patients showed reduced occupancy and stability of globally coordinated states and increased engagement of subcortical-dominant states, indicating impaired global stability and disrupted cortical–subcortical balance. After ECT, clinical symptoms improved significantly, accompanied by increased occupancy of states characterized by higher activity within the default mode and frontoparietal networks, as well as a highly integrated whole-brain state. These changes suggest improved overall brain coordination and cortical engagement rather than full normalization of baseline abnormalities. In contrast, subcortical-dominant states persisted after treatment, consistent with trait-like features of TRD. Neurotransmitter mapping showed that treatment-responsive states were associated with serotonergic, dopaminergic, and glutamatergic systems, whereas persistent subcortical states were linked to noradrenergic and cholinergic markers. Overall, these findings suggest that ECT improves symptoms by shifting brain dynamics toward more adaptive and integrated states, rather than fully restoring baseline function. This dynamic perspective provides new insight into the neural mechanisms of TRD and its treatment.

Introduction

Treatment-resistant depression (TRD) is defined as an inadequate response to at least two antidepressant trials of sufficient dose and duration. Affecting up to 55% of patients with major depressive disorder (MDD), TRD remains a major clinical challenge, with remission rates below 15% [[1], [2], [3]]. Beyond persistent symptoms, TRD is associated with marked reductions in quality of life and a high risk of suicidal behavior (4.66 attempts per 100 patient-years), resulting in substantial social and economic burden [[4], [5], [6]]. Understanding the neurobiological basis of TRD and identifying effective treatment mechanisms are therefore critical priorities in psychiatric research.

Electroconvulsive therapy (ECT) is the most effective treatment for achieving rapid remission in TRD [3]. However, the neural mechanisms through which ECT alters brain function are not fully understood. Most neuroimaging studies have relied on static functional connectivity (FC) to examine these effects. Prior work has shown that ECT modulates limbic and limbic–prefrontal networks and restores connectivity within the default mode network (DMN) in treatment responders [7,8]. Increased FC between frontal regions and posterior DMN areas, such as the posterior cingulate cortex and angular gyrus, has been identified as a consistent imaging finding [9]. Our previous study similarly showed that ECT influences FC within the frontoparietal network (FPN), language network, and DMN, supporting improvements in emotional and cognitive function [10]. However, static FC assumes that brain activity remains stable over time, which may overlook short-lived and recurring brain states relevant to depressive symptoms.

The brain is a dynamic system that continuously changes to support cognitive and emotional processes. Increasing evidence suggests that ECT acts by altering these time-varying patterns of brain activity. For example, ECT has been shown to change the variability of intrinsic activity in regions such as the dorsal anterior cingulate cortex and precuneus [11], and to influence network dynamics by increasing overall modular structure while reducing excessive variability within the DMN and somatomotor network (SMN) [12]. In addition, studies of critical dynamics indicate that ECT shifts brain activity from a rigid state toward a more flexible and coordinated pattern [13]. Together, these findings suggest that ECT may work by moving the brain away from a fixed, maladaptive pattern toward greater flexibility.

Despite these advances, most studies of brain dynamics after ECT have used sliding-window methods. Although this approach has linked changes in state occupancy to symptom improvement [14], it depends on arbitrary window lengths and has limited ability to capture rapid transitions between discrete brain states. The Hidden Markov Model (HMM) offers a more suitable framework by identifying distinct brain states without temporal smoothing [15,16]. Unlike simple clustering, HMM computes the precise probability of state activation at each time point, offering a more sensitive perspective into the microstate transitions associated with mental disorders, such as schizophrenia and bipolar disorder [17,18]. This approach has been used to identify state-related and trait-related features in MDD, including a tendency to spend more time in weakly connected states [19]. However, its application to longitudinal treatment studies remains limited. In particular, it is unclear whether ECT-related recovery reflects changes in specific brain states or improvements in transitions between them.

In addition to describing dynamic patterns, an important challenge is to understand the neurochemical basis of these changes. The monoamine hypothesis suggests that ECT affects dopaminergic, serotonergic, and noradrenergic systems, while more recent work highlights its role in restoring the balance between inhibitory (GABAergic) and excitatory (glutamatergic) activity [20,21]. Advances in multimodal analysis, such as the JuSpace framework [22], now allow integration of neuroimaging data with positron emission tomography (PET)-based maps of neurotransmitter systems. Examining the spatial relationships between HMM-derived brain states and neurotransmitter distributions may provide insight into how ECT influences brain activity at the molecular level.

In the present study, we applied a Gaussian HMM to resting-state functional MRI (rs-fMRI) data from patients with TRD before and after ECT, along with healthy controls (HCs). We aimed to characterize brain state dynamics using fractional occupancy (FO), averaged lifetime (ALT), and switching rate (SR). We also examined their associations with clinical improvement and neurotransmitter distributions. We hypothesized that [1]: TRD patients would show abnormalities in brain states related to cognitive and emotional processing at baseline [2]; ECT would shift state occupancy toward a more typical pattern [3]; the averaged functional activity (AFA) of each state would be related to specific neurotransmitter systems; and [4] improvements in depressive and anxiety symptoms would be associated with changes in state occupancy and stability over time.

Methods

Subjects and clinical assessments

Between June 2018 and December 2023, 74 patients with TRD were recruited from the Inner Mongolia Autonomous Region Mental Health Center, along with 66 age- and sex-matched HCs from the local community. The study was approved by the local Ethics Committee, and all participants provided written informed consent.

Inclusion criteria were: age 18–55 years, ≥9 years of education, and right-handedness. Exclusion criteria included a history of neurological disorders, major physical illness, substance abuse (alcohol, nicotine, or drugs), pregnancy or lactation, and MRI contraindications. TRD patients met DSM-IV criteria for MDD and were experiencing a current depressive episode confirmed by two psychiatrists. Clinical severity required a Hamilton Depression Rating Scale (HAMD-17) score ≥21. TRD was defined as <20% symptom improvement despite stable antidepressant treatment before enrollment. Patients with contraindications to ECT were excluded. HCs had no personal or first-degree family history of psychiatric disorders.

All participants completed clinical assessments and rs-fMRI at baseline. TRD patients were reassessed after completing ECT. All evaluators received standardized training to ensure consistency. Symptom severity was assessed using HAMD, the Hamilton Anxiety Rating Scale (HAMA), and the Brief Psychiatric Rating Scale (BPRS). Participants also completed self-report scales, including the Self-Rating Depression Scale (SDS), Self-Rating Anxiety Scale (SAS), and Automatic Thoughts Questionnaire (ATQ).

Intervention

TRD patients received ECT using the spECTrum 5000Q system. A standard course included 6–12 sessions, starting with three consecutive daily sessions, followed by sessions every other day (excluding weekends). After the initial six sessions, treatment was continued at the discretion of the attending psychiatrist if patients did not achieve sufficient relief of depressive symptoms. The total number of ECT sessions did not exceed 12. Detailed ECT and anesthesia protocols are described in the Supplemental Methods.

All patients continued their stable medication regimens during the study. Drug doses were converted into standardized equivalents using fluoxetine, alprazolam, and chlorpromazine as references [23,24]. Detailed medication information is provided in Table S1.

Imaging data acquisition and preprocessing

Imaging data were acquired on a 3.0 T S Skyra scanner using a gradient-recalled echo-planar imaging (EPI) sequence. Detailed parameters are provided in the Supplemental Methods. For TRD patients, scans were performed within 24 h before the first ECT session and within 24 h after the final session. During scanning, participants were instructed to rest with eyes closed, remain still, stay awake, and avoid structured thinking.

Rs-fMRI data were preprocessed using SPM12 and the RESTplus toolkit [25], implemented in MATLAB R2018b (http://www.mathworks.com/). Preprocessing steps (see Supplemental Methods) included slice timing correction, realignment, removal of nuisance signals, etc. Participants with excessive head motion (>2.5 mm translation or >2.5° rotation) were excluded.

Time series were extracted using the Schaefer atlas (100 cortical regions) and the Tian 2020 atlas (16 subcortical regions). The subcortical regions included the bilateral hippocampus (HIP), amygdala (AMY), posterior and anterior thalamus (pTHA and aTHA), nucleus accumbens (NAc), globus pallidus (GP), putamen (PUT), and caudate (CAU); all 16 regions were assigned to the subcortical network. Data from all participants (63 HCs, 72 TRD at baseline, 52 TRD post-treatment) were combined into a single dataset (43,384 time points × 116 regions) and standardized (z-scored) for analysis (Fig. 1A).

Fig. 1.

Schematic overview of the study. (A) Data acquisition and preprocessing. Resting-state fMRI data were acquired from patients with treatment-resistant depression (TRD, n = 74) before and after a course of electroconvulsive therapy (ECT), and from healthy controls (HCs, n = 66). Regional time-series were extracted from 116 regions of interest (ROIs) using the Schaefer-100 cortical and Tian-2020 subcortical parcellation schemes. (B) Hidden Markov Model (HMM) analysis. A Gaussian HMM was applied to the combined ROI time-series. The optimal model was selected using the Expectation-Maximization algorithm based on model fit and interpretability. Each latent state was characterized by its average functional activity (AFA) and functional connectivity (FC). Three dynamic metrics were derived: fractional occupancy (FO), averaged lifetime (ALT), and switching rate (SR). (C) Group comparisons. Longitudinal (pre-vs. post-ECT) and cross-sectional (TRD vs. HC) differences in dynamic metrics were evaluated. (D) Correlations with clinical measures. Partial correlation analyses were conducted to examine the relationship between HMM metrics and clinical symptom improvement, as measured by scales. (E) Neurotransmitter mapping. State-specific AFA maps were spatially correlated with PET-derived normative maps of 19 neurotransmitter receptors/transporters using Spearman correlations with false discovery rate (FDR) correction to identify the neurochemical basis of brain dynamics.

Hidden markov model

A Gaussian HMM was implemented using the hmmlearn library in Python, with parameters estimated via the Expectation–Maximization algorithm (maximum 1000 iterations). The number of states was tested from 5 to 12, with each model repeated at least five times to ensure stability. The final model (k = 6) was selected based on model fit and interpretability, as additional states showed minimal occupancy. Each state was defined by a mean vector (representing AFA) and a covariance matrix (representing FC between regions). From the model, three temporal metrics were derived: FO, ALT, and SR. FO reflects the proportion of time participants spent in each state. ALT, which represents the average number of time-points per visit to that state before switching to any other state, indicating the stability of a state [26]. SR can be computed by the rate of transitioning between all states, which represents the stability of brain dynamics per participant [27] (Fig. 1B).

Statistical analyses

Demographic and clinical differences between the TRD and HC groups were evaluated using Chi-square tests or Mann–Whitney U tests, as appropriate. Within the TRD cohort, longitudinal changes in clinical characteristics (pre-vs. post-ECT) were assessed using paired t-tests or Wilcoxon signed-rank tests, depending on the normality of the data distribution. All statistical analyses were conducted using SPSS version 25.0.

HMM metrics were transformed using inverse normal transformation to improve normality [19]. Group differences in HMM metrics across conditions (HC, TRD baseline, and TRD post-ECT) were evaluated using linear mixed-effects models (LMM), with condition as a fixed effect and subject ID as a random effect to account for intra-individual repeated measures (Fig. 1C). Age, sex, and education were included as covariates. Post hoc comparisons were performed using Wald tests, and multiple comparisons were corrected using the Bonferroni method (p < 0.05).

Correlation analyses between HMM metrics (FO, ALT, and SR) and clinical measures were performed using SPSS. Specifically, partial correlation analyses were conducted to evaluate the relationship between these HMM parameters and clinical scale scores (HAMD, HAMA, SDS, SAS, ATQ, and BPRS) (Fig. 1D). To evaluate intra-individual treatment relationships, partial correlation analyses were performed to [1]: explore whether baseline HMM metrics predict clinical reduction rates ([pre – post]/pre), and [2] examine the relationship between clinical reduction rates and longitudinal changes in HMM metrics (post - pre). In all analyses, age, sex, and years of education were included as covariates to control for potential confounding effects.

Spatial correlation with neurotransmitter

To examine neurochemical associations, we correlated AFA maps of each state with PET-derived maps of 19 neurotransmitter receptors and transporters from over 1200 healthy individuals [28]. These PET-derived maps included nine distinct neurotransmitter systems: dopamine (D1, D2, and DAT), serotonin (5-HT1a, 5-HT1b, 5-HT2a, 5-HT4, 5-HT6, and 5-HTT), glutamate (mGluR5 and NMDA), acetylcholine (α4β2, M1, and VAChT), histamine (H3), opioids (MOR), norepinephrine (NET), gamma-aminobutyric acid (GABAa), and cannabinoids (CB1). Analyses were performed using the JuSpace toolbox (https://github.com/juryxy/JuSpace). Spearman correlations were computed between state-specific AFA and neurotransmitter densities across cortical regions (Fig. 1E). False discovery rate (FDR) correction was applied to control for multiple comparisons.

Results

Demographic and clinical characteristics

A total of 74 patients with TRD and 66 HCs were initially recruited. Ten participants were excluded due to excessive head motion (two at baseline, five post-ECT, and three HCs), resulting in a final sample of 72 TRD patients at baseline, 63 HCs, and 52 patients who completed follow-up.

As shown in Table S2, there were no significant differences in age, sex, or education across groups at baseline (all p > 0.05). Baseline clinical and demographic characteristics did not differ between follow-up completers and non-completers (Table S3) or between the full TRD cohort and the follow-up subset (Table S4). Following ECT, patients showed marked clinical improvement, with significant reductions in HAMD, HAMA, SDS, SAS, ATQ, and BPRS scores (all p < 0.001; Table S5).

Spatiotemporal state dynamics and transition patterns

The HMM identified six distinct brain states, each defined by specific patterns of AFA (Table S6; Fig. 2) and FC (Table S7–S8; Fig. S1).

Fig. 2.

Overview of the averaged functional activity of the states. Left: Brain maps illustrating the averaged functional activity of regions in each state, grouped by eight functional networks. Right: Heatmap showing the averaged activity of 116 brain regions, organized according to their corresponding networks. Each row corresponds to one of the six spatial states, with values consistent with those shown in the brain maps. DAN = Dorsal Attention Network; DMN = Default Mode Network; FPN = Frontoparietal Network; LN = Limbic Network; SMN = Somatomotor Network; SN = Salience Network; VN = Visual Network.

State 1 showed relatively high FC within and between all networks (second only to State 3), with reduced AFA in the dorsal attention network (DAN) and subcortical regions. State 2 was characterized by moderate FC (weaker than States 1/3, stronger than States 4–6) and elevated AFA in the DMN and FPN. State 3 exhibited the highest FC across all networks, alongside reduced AFA in the DMN, FPN, and visual network (VN). State 4 showed intermediate FC within DAN (higher than State 5 but lower than others), with increased AFA in the DMN and limbic network (LN) and reduced AFA in the SMN. State 5 was marked by relatively stronger FC within the subcortex compared to States 4/6 and reduced AFA in the LN and VN. State 6 showed generally weak FC within and between all networks (exceeding only State 4) but increased AFA in the DAN, SMN, and VN.

State transitions were non-random and non-sequential (Fig. S2), with distinct transition patterns across HCs, baseline TRD, and post-ECT TRD groups.

Group differences in temporal dynamics

Significant group differences were observed in state-specific temporal metrics (Fig. 3; Fig. S3). For FO, HCs showed higher values than both baseline and post-ECT TRD patients in State 1 (p = 0.010 and p < 0.001, respectively). In contrast, post-ECT TRD patients showed higher FO in State 2 compared to HCs (p = 0.004). For State 3, HCs exhibited higher FO than both TRD groups (p < 0.001 and p = 0.003), although FO significantly increased after ECT relative to baseline (p = 0.003). In State 5, TRD patients showed consistently higher FO than HCs at both time points (p = 0.016 and p < 0.028). For ALT, post-ECT TRD patients showed shorter duration in State 1 compared with both HCs and baseline (p = 0.009). No significant group differences were observed in SR (Fig. S4).

Fig. 3.

Group comparisons of the temporal features. The hidden Markov model (HMM) identified six spatial states, with brain maps showing the mean functional activity for each state (blue to red indicates negative to positive values; range −2.78 to 2.78). This figure presents fractional occupancy and averaged lifetime for states 1, 2, 3, and 5; results for states 4 and 6 are provided in Supplementary Fig. S1. Positive values reflect increased neural activity, whereas negative values indicate decreased activity. Functional activity represents the averaged blood-oxygen-level-dependent (BOLD) time series within each state for each region. Violin plots depict group comparisons of temporal features (fractional occupancy and mean lifetime), with p-values shown above each comparison after Bonferroni correction. TRD: T1 = Treatment-resistant Depression At Baseline (N = 72); TRD: T2 = Treatment-resistant Depression After Completion of Follow-up Electroconvulsive Therapy (N = 52); HC = Healthy Controls (N = 63); DAN = Dorsal Attention Network; DMN = Default Mode Network; FPN = Frontoparietal Network; LN = Limbic Network; SN = Salience Network; VN = Visual Network.

Associations with clinical measures

State 6 FO was negatively correlated with SAS scores (r = −0.283, p = 0.037). In addition, reductions in State 4 FO and State 6 ALT were associated with the reduction rate of SDS and SAS, respectively (r = −0.308, p = 0.031; r = −0.328, p = 0.021; Fig. S5). Analysis assessing baseline predictors of treatment response revealed nominal, uncorrected correlations between HAMA reduction rates and baseline ALT of State 2 (r = 0.414, p = 0.017), baseline ALT of State 6 (r = −0.340, p = 0.034), and baseline FO of State 6 (r = −0.321, p = 0.024; Fig. S6).

Spatial correlation with neurotransmitter

State-specific AFA showed distinct relationships with neurotransmitter systems (Fig. 4). AFA in State 2 was positively correlated with 5-HT2a (r = 0.363, p < 0.001), D1 (r = 0.376, p < 0.001), GABAa (r = 0.360, p < 0.001), M1 (r = 0.328, p < 0.001), and NMDA (r = 0.321, p < 0.001) receptor densities. In contrast, State 3 AFA was negatively correlated with 5-HT2a (r = −0.603, p < 0.001), GABAa (r = −0.618, p < 0.001), and M1 (r = −0.519, p < 0.001). State 5 AFA showed positive associations with NET (r = 0.620, p = 0.002) and VAChT (r = 0.512, p = 0.002) distributions.

Fig. 4.

Spatial correlation between the averaged functional activity of the states and neurotransmitters (∗ = 0.05, ∗∗ = 0.01, ∗∗∗ = 0.001). 5HT1a = 5-HT subtype 1a receptor; 5HT1b = 5-HT subtype 1 b receptor; 5HT2a = 5-HT subtype 2a receptor; 5HT4 = 5-HT subtype 4 receptor; 5HT6 = 5-HT subtype 6 receptor; 5-HTT = serotonin transporter; α4β2 = nicotinic acetylcholine receptor subtype α4β2; CB1 = cannabinoid receptor; D1 = dopamine D1 receptor; D2 = dopamine D2 receptor; DAT = dopamine transporter; GABAa = GABA A receptor subtype; H3 = histamine receptor subtype H3; M1 = muscarinic acetylcholine receptor subtype 1; mGluR5 = metabotropic glutamate receptor 5; MOR = mu-opioid receptor; NET = norepinephrine transporter; NMDA = N-methyl-d-aspartate receptor; VAChT = vesicular acetylcholine transporter.

Discussion

In this study, we combined HMM with neurotransmitter atlas mapping to examine the spatiotemporal brain dynamics of TRD and their modulation by ECT. We identified six distinct brain states that can be broadly grouped into two dimensions: (i) persistent, trait-like abnormalities and (ii) treatment-related dynamic changes associated with clinical improvement. This approach clarifies how ECT works, suggesting that symptom relief comes from shifting the brain toward more adaptive and integrated modes, rather than a full reversal of baseline deficits.

At the trait level, TRD was marked by deficits in global stability and cortical–subcortical balance. One key feature was reduced engagement and stability of a globally coordinated baseline state (State 1), characterized by relatively high FC but reduced AFA in the DAN and subcortical regions. In healthy individuals, this pattern may reflect a stable baseline mode that maintains global coordination while limiting subcortical noise and external attentional demands [29]. Reduced involvement of the DAN—including the superior parietal lobule and frontal eye fields, which support top-down attention [30,31]—suggests a diminished ability to maintain stable, internally regulated brain activity. Consistent with this interpretation, TRD patients showed lower FO and shorter ALT in State 1, indicating reduced engagement and stability [32]. These findings align with prior evidence of DAN dysfunction in MDD, including reduced nodal efficiency and regional homogeneity [33,34]. In addition, decreased AFA in subcortical regions such as the basal ganglia, THA, and HIP is notable given the well-established structural and functional abnormalities in these regions in depression [35,36]. Together, these results suggest a reduced capacity to regulate abnormal activity in subcortical affective systems.

This instability was accompanied by persistent engagement of a subcortically dominated state (State 5), marked by increased subcortical connectivity and reduced AFA in the LN and VN. This pattern suggests a functional imbalance, where subcortical regions are strongly connected but less active, potentially reflecting weakened cortical control over core emotional systems. The LN—including the AMY, HIP, and cingulate cortex—is central to emotional processing [37,38], and its dysfunction has been linked to illness duration and executive deficits in depression [39]. Although previous studies suggest that ECT can enhance emotional regulation by increasing LN efficiency and connectivity [40], our findings indicate continued reliance on this subcortical-dominant state in TRD. Importantly, the persistence of reduced State 1 engagement and increased State 5 involvement after ECT suggests that these features represent trait-like vulnerabilities that are not fully corrected by acute treatment [19]. Such enduring abnormalities may contribute to the high relapse risk in TRD [41,42].

In contrast, clinical improvement was mainly linked to treatment-related changes in cortical engagement and global coordination. After ECT, patients showed increased occupancy of a state with elevated AFA in the DMN and FPN (State 2). These networks support self-related processing, emotional regulation, and cognitive control [43,44]. In MDD, the DMN is often associated with rumination [45,46], while the FPN typically shows reduced FC, indicating impaired cognitive control [47,48]. Increased engagement of this state after ECT likely reflects recovery of cortical activity and top-down regulation, helping to reduce cognitive and emotional symptoms.

ECT was also associated with increased occurrence of a highly integrated state (State 3), which showed the strongest FC across the brain, along with relatively lower AFA in the DMN, FPN, and VN. This pattern—strong global connectivity with lower regional activity—may indicate more efficient information processing, where broad coordination is achieved with less local demand. Changes in the VN are particularly relevant, as structural and functional visual system abnormalities have been linked to emotional processing in depression [49]. Recent evidence suggests that the visual system in MDD may become more self-contained and less influenced by higher-order control networks such as the DAN [50]. The recovery of this highly integrated state suggests that ECT enhances the brain's ability to engage in coordinated whole-brain activity, which may support adaptive cognitive and emotional function.

Neurotransmitter mapping further clarified the biological basis of these changes. Increased AFA in the DMN/FPN-dominant state (State 2) was positively associated with the distribution of 5-HT2a, D1, and NMDA receptors, consistent with known effects of ECT on monoaminergic and glutamatergic systems. These systems are key to synaptic plasticity, suggesting that ECT may promote a state of increased neural responsiveness that supports circuit-level changes [51]. In contrast, AFA in the highly integrated state (State 3) was negatively associated with 5-HT2a and GABAa receptors. Occipital GABA levels are known to be reduced in TRD [52], and successful recovery—particularly following ECT—has been associated with increased GABA concentrations and changes in 5-HT2a receptor binding in frontal and occipital regions [53,54]. These findings suggest that imbalances in inhibitory systems may limit the ability to engage this state at baseline, with ECT helping to restore this capacity.

Supporting a neurochemical basis for persistent dysfunction, State 5 AFA showed positive associations with NET and VAChT distributions. The noradrenergic system plays a key role in depression, with NET as an important treatment target [55], while VAChT regulates acetylcholine release, which is linked to mood and anxiety symptoms [56]. Chronic stress can increase hippocampal acetylcholine levels, leading to maladaptive synaptic changes [57]. These findings suggest that dysregulated noradrenergic and cholinergic systems may sustain this subcortical-dominant state, reinforcing its role as a marker of treatment resistance.

Taken together, these findings support a model in which ECT induces a functional “reset” of brain dynamics. Rather than fully correcting trait-like abnormalities in baseline stability and cortical–subcortical balance, ECT appears to compensate for these deficits by increasing the availability of more adaptive brain states. Greater cortical engagement (State 2) and improved whole-brain coordination (State 3) may help the brain work around persistent structural or functional limitations, supporting more flexible and efficient information processing. This model helps explain how significant clinical improvement can occur despite incomplete normalization of underlying dysfunction and highlights dynamic change as a key mechanism of treatment response.

Finally, two additional states (States 4 and 6), reflecting internally and externally oriented processing, showed limited involvement in treatment effects. State 4, with increased AFA in the DMN and LN and reduced sensorimotor activity, likely reflects an internally focused, ruminative pattern linked to psychomotor slowing [58,59]. State 6, characterized by higher AFA in the DAN, SMN, and VN, may reflect external attention and sensory engagement. Although some associations with symptom changes were observed, the lack of clear longitudinal changes suggests that these states play a smaller role in treatment-related dynamics and may instead represent relatively stable features of TRD.

Several limitations should be noted. First, patients remained on stable antidepressant medications, and drug effects on brain dynamics cannot be fully excluded. Second, sample attrition reduced the longitudinal dataset, although baseline similarities suggest limited bias. Third, the association analysis relied on normative, group-level receptor atlases rather than patient-specific imaging, which may not fully capture the individual neurochemical variations or disease-specific alterations inherent to our TRD cohort. Fourth, because the control group was scanned only at baseline in the present study, a repeated-measures comparison incorporating longitudinal changes in HCs was absent; thus, we cannot entirely attribute the observed pre-to-post changes solely to the ECT intervention. However, previous studies [60] have demonstrated that resting-state fMRI metrics exhibit high test-retest reliability and relative temporal stability in healthy individuals, suggesting that intra-individual baseline-to-follow-up fluctuations in healthy controls are typically limited compared to treatment-induced shifts. Finally, selecting the number of HMM states remains a methodological challenge [61,62], and different model choices may yield different results.

Conclusion

This study provides a dynamic perspective on TRD and its treatment with ECT based on HMM. TRD is characterized by persistent deficits in baseline stability and cortical–subcortical regulation, while clinical improvement is linked to increased cortical engagement and improved whole-brain coordination. These findings highlight the importance of viewing brain function as a dynamic process and suggest that effective treatments may work by reshaping, rather than fully normalizing, brain activity patterns.

Supplement Description

Supplement Methods, Results, Figure S1-S6, Tables S1-S8.appsec1

Author contributions

TL and LZ: performed data curation, conducted formal analysis, and contributed to the writing – original draft preparation as well as reviewing and editing. YH, ZG, PY, LC, YY, JZ, and YO: were responsible for data curation and conceptualization. DL and WG: contributed to methodology, writing – review and editing, and funding acquisition. All authors have read and approved the final version of the manuscript.

Fundings

This work was supported by the National Natural Science Foundation of China (82171508) and the Natural Science Foundation of Hunan (2025JJ60706).

Conflicts of interest

The authors declare no competing financial or non-financial interests.

Appendix A. Supplementary data

The following is/are the supplementary data to this article.