Distinct temporal neural dynamics of working memory in depressive patients: early encoding deficits and late maintenance correlates of rumination

Abstract

Working memory (WM) deficits and maladaptive rumination are hallmark features of major depressive disorder (MDD), yet the stage-specific electrophysiological mechanisms remain poorly understood. This study investigated neural dynamics during WM encoding and maintenance in individuals with depressive episode (DE) and their associations with rumination subfactors. A total of 59 DEs and 49 healthy controls (HCs) performed an n-back task (0-back and 2-back) while 64-channel electroencephalography (EEG) was recorded. Behavioral results demonstrated that the DEs exhibited impaired WM updating, characterized by significantly lower accuracy and slower reaction time (RT) under high cognitive load. Electrophysiological analyses revealed that during encoding, the DEs showed attenuated load-related modulation of the frontal N2 amplitude. During maintenance, the DEs demonstrated pathologically elevated occipital alpha power. Crucially, a significant group × alpha interaction (Wald χ² (1) = 9.258, p = 0.002) in predicting RT was observed. Higher alpha power was associated with slower RT in HCs (β = 0.006, p = 0.021), whereas this neurobehavioral coupling was disrupted and paradoxically reversed in the DEs (β = −0.003, p = 0.063). Within the DE group, alpha power was significantly negatively associated with brooding (β = −0.962, p = 0.006) and reflective pondering (β = −0.692, p = 0.035) independent of task load. Our findings reveal that impaired resource mobilization during encoding and disorganized functional inhibition during maintenance converge to drive WM deficits in depressive patients, with the latter uniquely tied to maladaptive rumination.

Introduction

Major depressive disorder (MDD) is a severe global health concern that contributes substantially to disability worldwide [1, 2]. Depression is marked not only by affective disturbances but also by prominent deficits in cognitive functioning [3]. Across acute and remitted phases, more than half of individuals with MDD have been reported to exhibit cognitive impairments, particularly in processing speed, visual selective attention, working memory (WM) and verbal learning [1, 4]. WM functions as a central hub for supporting goal-directed cognitive activities such as reasoning and problem solving [5, 6]. Previous studies have demonstrated a significant positive association between WM and general cognitive function [7]. Consequently, impairments in WM lead to adverse psychosocial outcomes [8] and substantially hinder the ability to carry out many everyday activities [9]. In addition, impairments in WM function have been linked to antidepressant treatment response [10, 11].

Increasing evidence suggests that WM deficits in MDD are closely related to emotion regulation, which may weaken individuals’ ability to manage negative emotions and heighten maladaptive responses to adverse experiences [12]. Psychological models have hypothesized links between WM and ruminative thinking to explain depressive symptoms from a cognitive–emotional perspective [13, 14]. According to the WM model [15], depressed mood narrows attentional scope and impairs the updating of WM. This makes it difficult for new, positive, or goal-relevant information to replace negative material, which in turn promotes rumination. In an alternative inhibition model, Linville [16] and Joormann [17] suggest that rumination arises from reduced inhibitory control, which hampers the removal of negative, irrelevant content from WM. Several randomized controlled trials (RCTs) have likewise demonstrated, at the behavioral level, a close association between WM capacity and rumination. In depression, RCTs have shown that incorporating direct WM training based on the Paced Auditory Serial Addition Task into standard treatment protocols results in greater reductions in both rumination and depressive symptoms in the combined intervention group, with improvements in depressive symptoms occurring secondary to decreases in rumination [18, 19]. Among healthy university students with high ruminative tendencies, WM-based cognitive control training similarly reduced rumination, with greater improvements in WM predicting larger decreases in ruminative thinking under subsequent naturalistic stress [20]. These findings underscore the potential of WM function as both a diagnostic biomarker and a therapeutic target for MDD. Hence, elucidating the altered WM processes and their underlying neurobiological mechanisms in MDD holds substantial clinical and theoretical significance. Although behavioral and theoretical models that emphasize WM dysfunction as a mechanism linking depression to rumination are well established, far less is known about their underlying neural substrates.

WM relies on temporally distinct stages, such as encoding and maintenance, which engage different neural activities [21]. Specific event-related potentials (ERPs) and event-related spectral perturbations (ERSPs) correspond to distinct neural functions, reflecting multiple information processing pathways. Accordingly, numerous studies have examined neural characteristics of WM in depression and how these alterations relate to emotional symptoms. In the ERP domain, patients with MDD typically exhibit intact whole-brain N1 and P2 responses, reduced occipital N2 amplitudes, and enhanced frontal P3 and central N4 amplitudes [22–24]. In the ERSP domain, MDD is characterized by reduced or unchanged frontal theta power and increased occipital alpha activity during the WM encoding phase [25, 26], as well as decreased frontal theta, occipital gamma, and occipital alpha oscillations during the maintenance phase [25–27]. Notably, alpha and gamma oscillatory reductions during the maintenance period are correlated with depression severity [26]. However, despite these advances, existing studies rarely characterize WM processing across its full temporal sequence involving encoding, manipulation, and maintenance within the same experimental framework. More importantly, while theoretical models link WM-related electrophysiological abnormalities to depressive symptoms and rumination, the underlying mechanisms and supporting biological evidence remain elusive.

To answer the above questions, we utilized EEG recordings from the n-back paradigm together with standardized clinical scale evaluations to detect and characterize WM-related neural features. The study was designed with two central objectives: (1) to systematically examine the temporal electrophysiological abnormalities of WM in depression through analyses of ERPs and ERSPs, including event-related synchronization (ERS) and event-related desynchronization (ERD), and (2) to explore the potential relationships linking aberrant WM-related electrophysiological activity to depressive symptoms and rumination.

Methods and materials

Sample size calculation

The required sample size for a repeated measures design with one between subject factor (two groups) and one within subject factor (two task loads) was estimated using G*Power 3.1. Based on Nikolin et al. [25], a total of 22–30 participants would provide 80% power to detect a moderate effect size (f = 0.32) at a 5% significance level under the assumption that the correlation among repeated measures ranges from 0.3 to 0.5. Spearman correlations were computed to assess the associations between the significantly altered EEG metrics and questionnaire scores. The sample size of 82 participants afforded 80% statistical power (α = 0.05) to identify medium-effect correlations (|ρ| = 0.3), consistent with Cohen’s guidelines. Considering the data distribution, the Generalized Estimating Equations (GEE) model was ultimately employed instead of the initially planned repeated-measures Analysis of Variance (RMANOVA). Accordingly, the study aimed to recruit 41 patients with depression and 41 healthy controls (HCs).

Participants

Patients with depressive episode (DE) and HCs from the same geographic region were recruited, with the DEs enrolled at the Affiliated Brain Hospital of Nanjing Medical University between August 2021 and July 2025. Diagnoses were confirmed using the Structured Clinical Interview for Diagnostic and Statistical Manual of Mental Disorders-IV-Text Revision (DSM-IV-TR) Axis I Disorders (SCID-I), and clinical assessments were conducted by two experienced psychiatrists using the Mini International Neuropsychiatric Interview. All participants were right-handed, were aged 16–50 years, had normal or corrected-to-normal vision, and had completed at least 9 years of education. Inclusion criteria for DEs were as follows: (1) a diagnosis of MDD according to DSM-IV-TR; (2) a 17-item Hamilton Depression Rating Scale (HAMD-17) score ≥7; (3) absence of other neurological or psychiatric disorders; (4) no history of electroconvulsive therapy in the previous six months; and (5) no substance dependence or abuse within six months. Criteria for HCs were identical, except for the requirement of no psychiatric history or MDD diagnosis and no current use of neurological or psychiatric medications.

A total of 60 DEs and 49 HCs were initially recruited. One patient was excluded due to a revised diagnosis of bipolar disorder, resulting in a final sample of 59 DEs and 49 HCs. The study was approved by the Ethics Committee of Nanjing Brain Hospital Medical Research (No.2017-KY076 and 2022-KY009-01), was conducted in accordance with the Declaration of Helsinki, and written informed consent was obtained from all participants prior to study initiation.

Clinical assessment

Severity of depressive symptoms in the DEs group was evaluated with the HAMD-17. The self-reported Beck Depression Inventory (BDI) was also administered [28]. In the current dataset, the BDI showed excellent internal consistency (Cronbach’s α = 0.954). In addition, we used the Ruminative Responses Scale (RRS) [29] to assess individuals’ ruminative thinking, including three factors: symptom rumination (SR), reflective pondering (RP) and brooding. The three rumination subcomponents also exhibited robust reliability, with Cronbach’s α values of 0.951 for SR, 0.874 for RP, and 0.794 for brooding.

Stimuli and task

WM performance was assessed using an n-back [30, 31] paradigm administered during EEG recording. The task comprised three blocks each of the 0-back and 2-back tasks, administered in a pseudo-randomized order. Participants were seated 30–50 cm from a computer screen, where white digits (0–9) were displayed on a black background for 500 ms with a 1500 ms interstimulus interval. Each block contained 40 trials, with 25% designated as targets, yielding a total of 6 blocks. The schematic representation of the task paradigm is shown in Fig. 1. Participants were asked to use their left and right index fingers to press the F and J keys for non-target and target cues, respectively. For the 0-back task, the target was a fixed digit (3), whereas for the 2-back task, the target was the digit presented two trials prior. Both reaction time (RT) and accuracy were recorded throughout the task. Before the formal experiment, participants reviewed the task instructions to ensure comprehension and completed one practice session for both the 0-back and 2-back tasks. The entire task lasted approximately 10 min, and accuracy and RTs were recorded automatically using E-Prime 2.0 software.

Fig. 1. Schematic illustration of the n-back task paradigm.

A In the 0-back task, participants identified whether the digit matched a pre-defined target (3). B In the 2-back task, they compared it with the digit two trials earlier. Each digit was presented for 500 ms with a 1500 ms interval. Participants pressed F for nontargets and J for targets and were instructed to respond as quickly and accurately as possible.

EEG recording and data preprocessing

In light of evidence that neural correlates of cognition in depression exhibit a diurnal pattern [32], all EEG data from depression participants were acquired during a fixed morning window (08:00–10:00) to control for this source of variability. The equipment was consistent with that used in our group’s prior publications. EEG signals were continuously recorded during the 0-back and 2-back tasks at a 1000 Hz sampling rate using a 64–channel Ag/AgCl electrode cap (Compumedics Neuvo) following the extended 10–20 system. The Quik-Cap Hydro Net (Australia) was employed, with the reference electrode positioned between Cz and CPz. Electrode impedances were maintained below 20 kΩ, and no online filtering was applied [32]. Preprocessing was performed in EEGLAB. Eight non-essential channels (F11–12, FT11–12, M1–2, CB1–2) were removed, and continuous data were bandpass filtered from 1–50 Hz with a 48–52 Hz notch filter. The sampling rate was downsampled to 250 Hz, and data were segmented into 3000 ms epochs (−1000 to 2000 ms relative to cue onset). Trials with incorrect responses were excluded, and signals were rereferenced to the average of all channels. Artifact rejection involved two steps: trials exceeding 100 μV were automatically inspected and removed, and bad channels were interpolated. Eye blinks, head movements, and muscle artifacts were eliminated using independent component analysis. The mean numbers of trials were 93.37 (0-back) and 80.83 (2-back) for DEs, and 99.12 (0-back) and 91.08 (2-back) for HCs.

Event-related potential analysis

Further analyses were conducted using the FieldTrip toolbox in MATLAB R2022b. ERPs for each participant and electrode were computed separately for the 0-back and 2-back tasks using a 300 ms baseline preceding the task cue (−400 to −100 ms). Component selection was informed by prior n-back ERP studies [25, 33, 34], grand-average waveforms, and topographical distributions. The N1 was averaged across three electrodes (O1, Oz, O2). The P1 was averaged across six electrodes (F1, Fz, F2, FC1, FCz, FC2), whereas the N2 was averaged over three electrodes (F1, Fz, F2). For each component, peaks were identified within predefined latency search windows (N1: 0–200 ms; P1: 100–300 ms; N2: 200–350 ms). Mean amplitudes were then computed within a narrow ±20 ms window centered on the individual peak latency [35].

Time-frequency analysis

Time-frequency analyses were performed across 2–45 Hz using a multi-taper approach within a −1000 to 2000 ms window relative to cue onset. A Hanning taper with frequency-dependent sliding windows (window length = 4/frequency) and a 20 ms step size was applied, allowing improved frequency resolution at lower frequencies and enhanced temporal resolution at higher frequencies. For each participant and task load (0-back/2-back), power values were averaged across trials. Oscillatory activities during encoding and manipulation were baseline-corrected using the decibel (dB) method relative to the −400 to −100 ms pre-stimulus period. In contrast, for the maintenance phase (1000–1500 ms post-stimulus), absolute power (µV²) was calculated without baseline correction. This approach was necessitated by the rapid succession of trials in the n-back design, which precluded a sufficiently uncontaminated baseline period for the late maintenance interval. Electrodes of interest (EOI) and frequency bands were determined based on previous literature and observed topographic distributions [26, 36]. To enhance statistical sensitivity and account for inter individual variability, subject specific peak centered time–frequency windows were used for each predefined electrode cluster. For each participant, the maximum or minimum power within the frequency and time range of interest was identified, and the mean power was then calculated within a small window centered on that peak. Specifically, for alpha power (10–12.5 Hz; O1, Oz, O2; 1000–1500 ms), power was averaged within a 100 ms (time) × 1 Hz (frequency) window around the individual maximum. For beta ERD (15–30 Hz; FCz, Cz; 300–900 ms), power was averaged within a 100 ms (time) × 4 Hz (frequency) window around the individual minimum. For gamma power (30–45 Hz; FCz, Cz; 1000–1500 ms), power was averaged within a 100 ms (time) × 8 Hz (frequency) window around the individual maximum, and for gamma ERD (30–45 Hz; FCz, Cz; 200–500 ms), within a 50 ms (time) × 8 Hz (frequency) window around the individual minimum. Because theta band (4–8 Hz) power is strongly influenced by phase-locked ERP components, it was excluded to ensure focus on non-phase-locked oscillatory activity [32, 37].

Statistical analysis

All statistical analyses were conducted using IBM SPSS Statistics 25.0 and R 4.2.2 software. Collectively, 4 BDI values were missing in the DE group, and RRS values were missing for 2 DEs and 1 HC. Group differences in age, years of education, BDI and RRS scores were examined using independent-samples t–tests or Mann–Whitney U tests, depending on the normality and homogeneity of variances. Chi-square tests were employed for sex comparisons. For each patient, a composite Medication Load Index (MLI) was calculated to quantify overall psychotropic medication burden [38, 39]. Detailed medication information for the patients is provided in Table S1. To control for its potential effects, we applied a data-preprocessing correction method. Specifically, each behavioral and neurophysiological outcome was regressed on the MLI using a linear model. The resulting regression coefficient was then used to compute median-adjusted residuals; these adjusted values constituted the final dataset for all group-level analyses. This approach statistically removes medication-related variance prior to the main analysis.

Given the study design, which involved repeated measures and covariates influencing WM, group (DE vs. HC) was treated as the between-subject factor, task load (0-back vs. 2-back) as the within-subject factor, and sex, age, and years of education as covariates. GEE models were applied to compare behavioral performance (accuracy, RT) and neurophysiological indices including ERP components (N1, P1, N2 amplitudes) and task-related oscillations (alpha, beta, gamma oscillations) between groups. To balance Type I and Type II errors, multiple comparison corrections were implemented hierarchically. No adjustment was applied across independent ERP/oscillatory components or for the main and interaction effects within each GEE model, as these represent distinct neurophysiological hypotheses [40]. False Discovery Rate (FDR) correction was strictly applied to post-hoc pairwise comparisons only when a significant interaction was observed.

To examine brain-behavior coupling across groups, GEE models were constructed to test the interaction between electrophysiological indices and group on task performance (accuracy and RT), controlling for task load, sex, age, and education. Within the DE group, GEE models (exchangeable correlation structure) were further applied to control for task loads to evaluate the associations between neural deficits and clinical measures (BDI and RRS subfactors).

Sensitivity analyses were performed to assess the robustness and consistency of the findings. For outcomes that did not reach statistical significance, Bayes factors (BF) were computed to evaluate the strength of evidence. In contrast to traditional frequentist approaches, Bayesian analysis enables quantitative assessment of the evidence in favor of the alternative hypothesis, the null hypothesis, or the insensitivity of the data [41, 42]. Overall, a BF less than 1/10 or between 1/10 and 1/3 indicates strong or moderate evidence, respectively, in support of the null hypothesis. Conversely, a BF greater than 10 or between 3 and 10 represents strong or moderate evidence favoring the alternative hypothesis. A BF value between 1/3 and 3, however, suggests that the data are insensitive for distinguishing between the competing hypotheses [41, 43]. Therefore, we further employed Bayesian RMANOVA and a Bayesian mixed model using the open-source software JASP (Version 0.97) to examine the results with p ≥ 0.05.

Results

Participant characteristics and clinical profiles

There was no significant group difference in sex distribution (χ² = 0.898, p = 0.343). Age and years of education were non-normally distributed in both groups (all Shapiro–Wilk p < 0.01). Therefore, Mann–Whitney U tests were conducted. The DE group differed significantly from HCs in age (U = 1128, p = 0.049) and years of education (U = 425, p < 0.001). For emotional measures, the DE group showed significantly higher scores across all self-reported and clinician-rated scales. Detailed descriptive and inferential statistics are presented in Table 1.

Table 1.

Demographic and clinical features and questionnaire scores of all participants.

Variables	Mean ± SD or Median (IQR)		χ²/U value	p value
	DE group (n = 59)	HC group (n = 49)
Sex (Male/ Female)	27/32	18/31	0.898a	0.343
Age	23.00 (9.00)	24.00 (3.00)	1128.000b	0.049*
Education years	15.00 (3.00)	17.00 (2.00)	425.000b	<0.001***
Medication load index	2.00 (1.00)	-	-	-
HAMD-17	20.35 ± 5.65	-	-	-
BDI	21.00 (18.00)	1 .00 (4.00)	181.500b	<0.001***
RRS	59.00 (18.00)	35.00 (16.00)	415.000b	<0.001***

DE depressive episode, HC healthy control, HAMD-17 17-item hamilton depression rating scale, BDI beck depression inventory, RRS ruminative responses scale.

^aχ² values were obtained using the Chi–square test.

^bU values were obtained using the Mann–Whitney U test.

*p < 0.05. **p < 0.01. ***p < 0.001.

Behavioral performance: impaired updating in depression

A GEE model revealed a significant group × task load interaction on both accuracy (Wald χ² (1) = 21.444, p < 0.001) and RT (Wald χ² (1) = 6.917, p = 0.009). Post-hoc comparisons showed that for accuracy: HC 0-back = DE 0-back > HC 2-back > DE 2-back; for RT: HC 0-back = DE 0-back < HC 2-back < DE 2-back. Detailed descriptive statistics are summarized in Table 2, indicating a specific deficit in WM updating under high cognitive demand in depression.

Table 2.

GEE results for working memory behavioral performance.

Variables	Median (IQR) back 2-back		Interaction effect Wald χ² (p)	Main effect Wald χ² (p)	Main effect Wald χ² (p)	Post-hoc
	DE group (n = 59)	HC group (n = 49)	Group × Task load	Group	Task load
Accuracy	0.97 (0.05) 0.77 (0.12)	0.98 (0.03) 0.86 (0.06)	21.444 (<0.001)***	33.334 (<0.001)***	343.414 (<0.001)***	HC0 = DE0 > HC2 > DE2
RT (ms)	434.06 (79.17) 693.98 (218.66)	413.50 (75.21) 606.78 (222.98)	6.917 (0.009)**	10.149 (0.001)**	319.207 (<0.001)***	HC0 = DE0 < HC2 < DE2

DE depressive episode, HC healthy control, RT reaction time, ms milliseconds.

*p < 0.05. **p < 0.01. ***p < 0.001.

Attenuated load-related N2 modulation during WM encoding in depression

As illustrated in Fig. 2, a significant group × task load interaction was found for N2 amplitude (Wald χ² (1) = 10.034, p = 0.002), which remained robust in a cluster-based permutation test without covariates (Fig. S1). Post-hoc analyses revealed a significant load effect in both groups. The 2-back task elicited less negative N2 amplitudes compared to the 0-back task (DE: estimate = 0.545, q = 0.0002; HC: estimate = 1.093, q = 0.0002). Importantly, the magnitude of load-related N2 modulation (0 → 2-back change) was substantially larger in HCs, nearly twice that observed in DEs. Between-group comparisons at each load level were not significant (0-back: estimate = 0.625, q = 0.134; 2-back: estimate = 0.077, q = 0.826). Early sensory components (N1, P1) showed no significant interaction or main effects (all p > 0.05, BF < 1). Topographic maps and violin plots for N1 and P1 are shown in Fig. S2.

Fig. 2. ERPs, topographic maps, and amplitude distributions.

A ERPs and topographic maps: (bottom) Grand-averaged ERP waveforms (mean ± SE) at FCz electrode, in which the shaded area indicates the N2 time window of interest; (upper) topographic maps for N2 in different groups and task loads. B Violin plots showing N2 amplitudes for each group across task loads. Boxplots display the median and IQR (whiskers = 1.5 × IQR); points denote individual participants, and grey lines indicate within-subject pairings. C Plot of interaction N2 amplitudes. DE depressive episode, HC healthy control.

Elevated occipital alpha oscillations during WM maintenance in depression

During the WM maintenance phase, for alpha power, the group × task load interaction was nonsignificant (Wald χ² (1) = 1.067, p = 0.302, BF = 0.315). In contrast, significant main effects of group (Wald χ² (1) = 4.645, p = 0.031; DE > HC) and task load (Wald χ² (1) = 7.240, p = 0.007; 0-back > 2-back) were observed (Fig. 3). For gamma power, no interaction or main group effect was detected (all p > 0.05, BF < 1), while a significant main task load effect (Wald χ² (1) = 13.243, p < 0.001; 0-back >2-back) was found. During the encoding and manipulation phase, neither significant group × task load interaction nor main effects were observed for beta ERD or gamma ERD (all p > 0.05, BF < 1). Similar results were found by a cluster-based permutation test without controlling covariates (see Fig. S3). The corresponding topographic maps, ERSPs, and violin plots for beta ERD and gamma activity are provided in Fig. S4.

Fig. 3. Topographic maps, ERSPs, and oscillatory activity distributions of alpha oscillations.

A Topographic distributions and ERSPs of alpha oscillations at the Oz electrode during WM maintenance for each group. B Violin plots of WM maintenance alpha power. Boxplots display the median and IQR (whiskers = 1.5 × IQR); points denote individual participants, and grey lines indicate within-subject pairings. Significant main effects are annotated on the plot. DE depressive episode, HC healthy control.

The complete statistical results for all electrophysiological indices are summarized in Table 3. Results based on raw values before MLI adjustment are provided in Supplementary Table S2 and were consistent with the adjusted analyses.

Table 3.

GEE for working memory–related ERPs and oscillations.

Variables	Median (IQR) 0-back 2-back		Interaction effect Wald χ² (p)	Main effect Wald χ² (p)	Main effect Wald χ² (p)
	DE group n = 59	HC group n = 49	Group × Task load	Group	Task load
N1 Amplitude (μV)	−2.45 (2.47) −2.38 (2.89)	−2.57 (4.21) −2.49 (4.40)	0.294 (0.588) BF = 0.240	1.098 (0.295) BF = 0.703	1.089 (0.297) BF = 0.233
P1 Amplitude (μV)	2.02 (2.21) 1.97 (1.84)	2.70 (2.38) 2.60 (2.59)	0.329 (0.567) BF = 0.253	2.471 (0.116) BF = 0.937	0.878 (0.349) BF = 0.215
N2 Amplitude (μV)	−1.48 (2.68) −0.97 (1.91)	−2.87 (2.44) −1.53 (2.53)	10.034 (0.002)** BF = 19.400	8.430 (0.004)** BF = 1.234	79.360 (< 0.001)*** BF = 1.283E+11
Alpha power (μV2) –WM maintenance	3.20 (10.54) 2.68 (7.85)	2.22 (5.10) 2.08 (4.32)	1.067 (0.302) BF = 0.315	4.645 (0.031)* BF = 0.946	7.240 (0.007)** BF = 3.844
Beta ERD (dB) –WM encoding	−2.38 (1.96) −2.74 (1.85)	−2.28 (2.24) −2.05 (2.07)	0.808 (0.369) BF = 0.306	0.783 (0.376) BF = 0.534	2.761 (0.097) BF = 0.588
Gamma ERD (dB) –WM encoding	−0.69 (0.83) −0.88 (0.72)	−0.75 (0.65) −0.64 (0.73)	1.318 (0.251) BF = 0.377	1.585 (0.208) BF = 0.690	0.164 (0.686) BF = 0.156
Gamma power (μV2) –WM maintenance	0.33 (0.40) 0.29 (0.36)	0.33 (0.29) 0.32 (0.23)	0.007 (0.934) BF = 0.197	1.589 (0.208) BF = 0.748	13.243 (< 0.001)*** BF = 53.715

DE depressive episode, HC healthy control.

*p < 0.05.**p < 0.01. ***p < 0.001.

Associations between electrophysiological indices and clinical symptoms, behavioral performance

As shown in Fig. 4, GEE analyses across the full sample revealed a significant group × alpha interaction in predicting RT (Wald χ² (1) = 9.258, p = 0.002). Simple slope analysis clarified that higher alpha power was significantly associated with slower RT in the HC group (β = 0.006, Wald χ² (1) = 5.340, p = 0.021), whereas this neurobehavioral coupling was disrupted in the DE group (β = −0.003, Wald χ² (1) = 3.449, p = 0.063). No significant interactions or main effects involving N2 amplitudes or accuracy were observed (p > 0.05). Additional DE-only GEE analyses showed that the alpha–behavior associations were not significantly moderated by rumination subfactors, including brooding and reflective pondering (all interaction p > 0.05; Supplementary Table S4).

Fig. 4. Association between alpha power and behavioral performance and rumination.

A Predicted RT as a function of alpha power in the DE and HC groups. A significant alpha × group interaction estimated from the GEE model was observed, indicating different brain–behavior coupling patterns between groups. B Scatterplot showing the relationship between alpha power and RT in each group. Regression lines represent linear model fits with 95% confidence intervals. C Relationship between alpha power and brooding in the DE group. Each point represents an observation under different task loads, and grey lines connect repeated measures within participants. D Relationship between alpha power and reflective pondering in the DE group.

Within the DE group, alpha power during the WM maintenance phase was significantly negatively associated with brooding (β = −0.962, Wald χ² (1) = 7.451, p = 0.006) and reflective pondering (β = −0.692, Wald χ² (1) = 4.441, p = 0.035). Notably, task load did not significantly moderate these clinical–neural associations (all p (scale × task load) >0.05).

No significant associations were found between N2 amplitudes and clinical scales, nor between alpha activity and BDI scores (all p > 0.05; see Supplementary Table S3 and Fig. S5).

Discussion

Our study demonstrates that depression is characterized by stage-specific electrophysiological abnormalities during WM that are uniquely linked to rumination. At the behavioral level, the diminished accuracy and prolonged RT under high load in DEs confirm a specific deficit in WM updating, providing a functional context for the observed neural alterations. During encoding, the attenuated load-related modulation of frontal N2 amplitude in DEs points to an initial failure in mobilizing cognitive resources. More critically, the correlation between alpha power and RT during the maintenance phase revealed that neurobehavioral coupling was disrupted and paradoxically reversed in patients, suggesting that alpha oscillations in depression no longer serve as an organized inhibitory gate but rather reflect a state of neural inefficiency. Furthermore, alpha oscillatory power was negatively correlated with specific ruminative cognitive styles rather than general symptom severity. Overall, our findings delineate a unique profile of WM impairment in depression, suggesting that cognitive deficits and maladaptive rumination share a common neural substrate.

Behaviorally, our GEE model revealed a significant group × task load interaction, with DE patients demonstrating lower accuracy and prolonged RT specifically in the 2-back task. This difficulty related performance decline, contrasting with the comparable 0-back performance across groups, indicates that WM abnormalities in depression are primarily driven by impairments in updating processes rather than basic attentional focus. The clinical relevance of these updating deficits is supported by evidence that MDD patients exhibit persistent WM impairments even during remission [44], and such deficits represent a prominent cognitive marker associated with suicidal behavior [45]. Our findings align with the study that identified updating as a core executive dysfunction in unipolar depression [46]. The observed behavioral cost suggests a fundamental deficit in the ability to rapidly replace outdated information with new task-relevant targets, a mechanism potentially linked to the persistence of irrelevant thoughts in depression [47]. Furthermore, impaired performance under high cognitive demand in DEs mirrors the updating biases frequently observed in studies using affective material [48]. Notably, recent research indicates that such deficits may be modifiable; virtual reality-based WM training has been shown to enhance frontotemporal connectivity and improve these specific behavioral indices [49]. These behavioral results provide a critical foundation for interpreting the subsequent electrophysiological outcomes.

During WM processing, the N2 component, originating from the anterior cingulate cortex (ACC) and its associated networks, reflects conflict monitoring and mismatch detection [50, 51]. In the present study, patients with depression demonstrated weaker load-related modulation of N2 amplitude (0-back vs. 2-back), rather than a simple amplitude reduction at either load level. This aligns with previous evidence that individuals with a history of depression exhibit altered ERP responses during demanding tasks, suggesting reduced neural efficiency and the use of distinct cognitive strategies even during remission [52]. Notably, this blunted sensitivity to task demands may represent a core deficit in neural flexibility. Recent evidence in other psychiatric populations, such as obsessive-compulsive disorder, has identified that load-dependent neural modulation defined as the difference between task loads is a more sensitive marker of cognitive dysfunction than activity at a single load level, and can even predict individual response to cognitive behavioral therapy [53]. Our findings extend this concept to MDD, suggesting that the inability to flexibly adjust N2-related control resources is a critical mechanistic index of the neural capacity under increasing cognitive demand. Such a deficit may be reversible through targeted interventions. For example, n-back WM training has been reported to enhance N2 amplitude during high demand tasks, accompanied by faster responding and improved suppression of irrelevant information [54]. Conversely, disrupting the prefrontal–ACC circuitry via brain stimulation has been shown to reduce N2 amplitude and impair WM [55]. While these studies primarily report changes in N2 amplitude rather than load dependent modulation, its causal significance as a potential therapeutic target is supported by intervention evidence. Therefore, while load related N2 modulation may not serve as a direct sensitive biomarker for overt behavioral performance, it remains a valuable mechanistic index for the dynamic recruitment of cognitive resources and a promising target for future clinical interventions.

Alpha oscillations, closely regulated by corticothalamic feedback loops, play a fundamental role in modulating cortical excitability and inhibiting task irrelevant regions [56–59]. In the current study, the DE group exhibited significantly higher alpha power in the occipital region during the task compared to the HC group. Although enhanced alpha activity has occasionally been interpreted as a compensatory mechanism [26], this view is not supported by our behavioral findings, where the DE group demonstrated impaired accuracy and prolonged RT. According to the alpha-BOLD anticorrelation theory, EEG alpha power is strongly and negatively correlated with cortical metabolic activity (BOLD signal), particularly within the occipital cortex [60]. Thus, the overall enhancement of alpha power in the DE group may reflect reduced cortical activation or insufficient task-related neural engagement, rather than effective compensation. Furthermore, GEE analyses revealed a significant group × alpha interaction in predicting RT, further elucidating the inter-group divergence in neurobehavioral coupling. In the HC group, alpha power was significantly and positively correlated with RT, meaning that lower alpha power was associated with faster response speeds. This aligns with the classic theory that reduced alpha power reflects increased cortical excitability, mediating more efficient information processing [60]. In contrast, the DEs exhibited a reversed coupling tendency, characterized by a marginal negative correlation between alpha power and RT. While this marginal significance may be attributed to the current sample size, the reversal in directionality suggests a distinct pathophysiological process in DEs. For these individuals, relatively lower alpha power (within the group) was associated with slower response times, which may reflect the “hijacking” of cognitive resources by internal negative interference, such as rumination [14, 61].

This inference is further supported by the significant negative correlation between alpha activity and ruminative subcomponents within the DE group. Rumination is essentially defined by ineffective cognitive control over negative self-referential information [62]. During the WM maintenance phase, although the DE group generally remained in a state of low cortical arousal (high alpha), intrusive ruminative thoughts may act as a potent hot cognitive interference [63, 64]. This forces a resource competition between task maintenance and internal emotional processing, leading to a disruption of the alpha-mediated gating mechanism that originally served to inhibit interference. The excessive consumption of limited internal resources then manifests as further prolongation of RT. In summary, the enhanced occipital alpha power in DEs reflects a general deficit in cognitive engagement and inefficient allocation of cortical resources. Our findings provide electrophysiological evidence for abnormal emotion-cognition coupling in depression, revealing how hot affective interference impairs cold cognitive functions by hijacking resources and destabilizing neural oscillations. Such maladaptive coupling may occupy working memory capacity, impair cognitive performance, and disrupt emotional regulation, thereby driving the progression of depressive pathology [65].

Additionally, one secondary finding emerged. Significant differences were observed only for the main effect of task load in gamma oscillations during the maintenance phase, with stronger activity in the 0-back task compared to the 2-back task. Previous intracranial EEG recordings in two epilepsy patients indicated that gamma activity (38 Hz) represents memory maintenance in the absence of external stimuli and increases with cognitive load [66], seemingly contradicting our findings. However, reviews have noted that gamma rhythms have been largely neglected in verbal WM studies, with no consistent evidence for sustained gamma activity during the delay period in EEG research overall [67]. Thus, the relationship between gamma oscillations and memory load during WM maintenance remains inconclusive. Contrary to expectations, gamma activity during the maintenance period did not escalate with increasing task load in our findings.

Limitations

Several limitations of this study should be acknowledged. First, the majority of depressive patients were inpatients receiving combined antidepressant and anxiolytic treatments. Although we statistically controlled for medication effects on neural oscillations using the MLI, evidence suggests that monoaminergic systems contribute to WM processing [68]. Future studies should aim to replicate the present findings in medication-free cohorts. Second, while our study preliminarily explored the potential links between WM processes, emotional regulation, and the pathophysiology of depression, the cross-sectional design precludes causal inferences. Longitudinal studies are therefore warranted to clarify the temporal and mechanistic relationships underlying these associations. Finally, the brain–symptom association analyses were conducted in a moderate sample without global multiplicity correction, which may limit statistical sensitivity. Future studies with larger samples are warranted to validate these findings and further clarify potential relationships.

Conclusion

The present study identifies stage-specific electrophysiological signatures of WM impairment in depression, characterized by attenuated load-dependent N2 modulation during encoding and pathologically elevated occipital alpha power during maintenance. Our findings demonstrate that while encoding deficits reflect impaired resource mobilization, the maintenance abnormalities are marked by a disrupted alpha-performance coupling and a specific association with rumination. In summary, these results provide a mechanistic account of how disorganized functional inhibition and clinical rumination converge to impair cognitive updating in depression. This evidence underscores the potential of targeting these distinct spatio-temporal neural dynamics to develop individualized interventions for ameliorating cognitive and affective symptoms in depression.

Author contributions

This study is the outcome of a collaborative work process. Contributions were as follows: Conception and design of the study: XW, ZY, HZ. Acquisition and formal analysis of data: HZ, XW, QL. Data collection: XW, HZ, SZ. Writing–original draft: XW, HZ. Writing–review & editing: ZY, QL, HZ, XW. Funding acquisition: ZY, QL. All authors contributed to the article and approved the submitted version.

Funding

This work was supported by the National Science Foundation of China (62571115, 82571749, 82271568, 82301718); the Jiangsu Medical Innovation Center for Mental Illness (CXZX202226); the Jiangsu Provincial Key Research and Development Program (BE2023667); Jiangsu Provincial Natural Science Youth Fund (BK20230154).

Data availability

The datasets generated during the current study are not publicly available due to the privacy, but are available from the corresponding author on reasonable request.

Competing interests

The authors declare that the research was conducted without any commercial or financial relationships that could be construed as a potential conflict of interest.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41398-026-04232-x.