Disrupted working memory event-related network dynamics in multiple sclerosis

Abstract

In multiple sclerosis (MS), working memory (WM) impairment can occur soon after disease onset and significantly affects the patient’s quality of life. Functional imaging research in MS aims to investigate the neurophysiological underpinnings of WM impairment. In this context, we utilize a data-driven technique, the time delay embedded-hidden Markov model, to extract spectrally defined functional networks in magnetoencephalographic (MEG) data acquired during a WM visual-verbal n-back task. Here, we show that the activation of two networks is altered in relapsing remitting-MS patients. First, the activation of an early theta prefrontal network linked to stimulus encoding and attentional control significantly decreases in MS compared to HC. This diminished activation correlates with reduced accuracy and higher reaction time, suggesting that impaired attention control impacts task performance in MS patients. Secondly, a frontoparietal network characterized by beta coupling is activated between 300 and 600 ms post-stimulus, resembling the event-related P300, a cognitive marker extensively explored in EEG studies. The activation of this network is amplified in patients treated with benzodiazepine, in line with the well-known benzodiazepine-induced beta enhancement. Altogether, the TDE-HMM technique extracts task-relevant functional networks showing disease-specific and treatment-related alterations, revealing potential new markers to assess and track WM impairment in MS.

Dynamic network analysis of MEG data reveals spatiospectral activation profiles underpinning working memory impairment in multiple sclerosis, distinguishing between the disease-induced and treatment-related neurophysiological alterations.

Introduction

Multiple sclerosis (MS) is the most common chronic neuroinflammatory disease of the central nervous system (CNS) in young adults¹. Demyelination and neurodegeneration of the CNS characterize the complex pathophysiology of this disease, resulting in a vast range of physical, neuropsychiatric, and cognitive symptoms^1,2. Evidence of cognitive impairment (CI) appears in about 50% of the people with MS (pwMS), with cognitive deficits predominantly in attention, working memory, or information processing speed^2–5. Working memory (WM) consists of encoding, maintaining, and retrieving a limited quantity of information for a few seconds or less⁶. Since these fundamental processes support all cognitive abilities, from language comprehension to mathematical reasoning, WM impairment majorly impacts the patients’ daily activities^2,3,5.

WM processes are performed via the synergetic activation of prefrontal, parietal, and temporal regions, forming large-scale brain networks that transiently activate throughout the task^7,8. Functional magnetic resonance imaging (fMRI) studies have described the spatial configuration of these dynamic networks and their altered activation in pwMS^5,9,10. However, the slow hemodynamic response captured by the fMRI signal overlooks the millisecond temporal evolution of the electrophysiological activity underlying cognitive processing¹¹. Additionally, recent studies have observed reduced fMRI signal quality in pwMS, suggesting that MS-induced hypoperfusion may alter cerebrovascular coupling¹².

Electrophysiological measurements, instead, sample brain activity with milliseconds temporal resolution and acquire the electric (via electroencephalography, EEG) or magnetic (via magnetoencephalography, MEG) fields resulting directly from the neuronal activity. Mainly studied in EEG data, the WM event-related response shows a positive peak 300 ms after stimulus onset (in the central frontal-parietal line) – the P300¹³. This peak is delayed and reduced in amplitude in pwMS, making the P300 a recognized cognitive marker for WM impairment in MS^14–16. Additionally, MEG time-frequency studies report reduced hippocampal theta and occipital alpha activity associated with worse task performance^16,17. This time-frequency description of MS-induced WM alterations focuses on region-specific activity. However, theta and alpha rhythms are known to lead to neural oscillatory synchronizations forming large-scale brain networks¹⁸.

The more recent MEG literature on MS presents a few dynamic functional connectivity studies, revealing the frequencies and mechanisms by which the transient network connections arise and how the MS-induced alterations are related to behavioural outcomes, i.e. fatigue^19–21. However, these studies are mostly based on resting-state (RS) data^19,20. Whereas the role of RS networks in cognition has long been established²², RS data may lack information on the temporal dynamics of domain-specific cognitive processes. Task data, on the other hand, allows us to link a functional network to WM-specific processes²³. In our previous exploratory analysis²⁴, we decomposed the WM brain dynamics in a healthy cohort, recorded by MEG during a visual-verbal n-back task, in transient functional networks that unfold over tens of milliseconds, hence, the temporal resolution of dynamic cognitive processes. While the spatial layout resembles the classical fMRI networks, we also explored the spectral content and event-related activation of each state, unveiling the potential role of each state in the neuropsychological understanding of WM^24,25.

We hypothesize that, by applying the same methodology on a dataset that also includes task MEG data of pwMS, the TDE-HMM model can help reveal the data-driven functional networks underpinning WM-specific processes, the activation of which is altered in pwMS, leading to a deeper understanding of the MS-induced WM impairment²³. In our results, we observe that MS reduces the activation of the early theta prefrontal network associated with stimulus encoding, suggesting that MS affects attentional control, directly impacting task performance. Additionally, we observe that treatments based on benzodiazepines have a remarkable effect on the network activation, amplifying the activation of the M300 beta frontoparietal state.

Results

Dataset composition

Table 1 shows the composition of the dataset we work with and reports the results on the demographics of the dataset. Two observations are worth mentioning. First, the BR1 and BR2 groups present the same proportion of HC, MS B−, and MSB+; however, BR1 comprises only female individuals. Second, the pwMS treated with benzodiazepine are only female patients. This led us to conduct a few analyses on subsets of the whole dataset, to test the reliability and robustness of the results.

Table 1.

Demographics

			Age mean (SD)	Education mean (SD)	Disease duration median (IQR)	EDSS median (IQR)	CVLT-II z-score data median (IQR)	SDMT z-score data median (IQR)
BR1	HC	Female (13)	44 ± 12	14 ± 2	NA	NA	0.13 (0.85)	−0.09 (0.85)
	MS B−	Female (19)	45 ± 10	14 ± 3	13 (14.3)	2.5 (1)	0.24 (1.23)	0.005.(1.35)
	MS B+	Female (7)	49 ± 7	13 ± 3	19(3.5)	3 (1.8)	−0.09 (1.96)	−0.18 (0.57)
BR2	HC	Female (10)	49 ± 11	15 ± 2	NA	NA	0.45 (1.29)	0.43 (1.23)
		Male (15)	50 ± 11	16 ± 2	NA	NA	−0.41 (1.48)	0.29 (1.56)
	MS B−	Female (17)	49 ± 9	16 ± 3	17(9.3)	2.5 (1.6)	0.02 (1.48)	0.005 (2.67)
		Male (18)	48 ± 9	14 ± 2	18 (12)	2.8 (1.5)	0.07 (2.25)	−0.14 (1.13)
	MS B+	Female (9)	13 ± 3	13 ± 3	11 (9)	3.5 (1.8)	0.021 (1.48)	−0.47 (0.92)
				p-value <0.05 Significant difference between HC, MS B−, and MS B+, only when considering the whole dataset.(one-way ANOVA test)		p-value <0.05 Significant difference between MS B+ and MS B− considering: • Whole dataset • Only female • Only BR2 (Wilkinson’s rank sum test)

The table presents the demographics of the datasets considering all the subdivisions: sex (Male and Female), disease and treatment conditions (HC, MS B−, and MS B+), and the scanner update (BR1 and BR2). We used a one-way ANOVA test with treatment status as the grouping variable to evaluate the difference in age, education level, and neuropsychological scores (for the two tests CVLT-II and SDMT) across the three groups (HC, MS B−, and MS B+) for the whole dataset, for only female, for BR2, and for BR1, separately. In these same subdivisions, we tested the difference in disease duration and EDSS score between MS groups with a Wilkinson’s rank sum test.

EDSS expanded disability status scale, CVLT-II California Verbal Learning Test – II, SDMT Symbol-to-Digit modality test, NA not applicable.

Starting from the whole group TDE-HMM inference, we investigated the group differences in ER activation profiles considering (1) the whole dataset and (2) sex-matched cohorts. In the main paper, we report the results on the HMM inference and group differences in ER activation for the whole dataset. Then, we identified a female cohort within the whole dataset; this female cohort is balanced in the number of subjects recorded with BR1 and BR2, and the three female groups of interest (HC, MS B−, and MS B+) are matched by age and education (Table 1). We also considered the male group (only from BR2), in which HC and MS B- are matched by age and education level, Table 1. The results of the analysis of the sex-matched group differences in ER activation profiles are reported in the Supplementary Materials, Supplementary Section 1, Supplementary Figs. S1–S4.

Lastly, to roll out a possible effect of the scanner update as a confound in our results, we ran the TDE-HMM on the BR1 and BR2, separately. The states’ description and group difference analyses of these inferences are reported in the Supplementary Materials, Supplementary Section 2, Supplementary Figs. S6 and S7. All the inferences report task-relevant states with the same spatio-spectral features and the same MS-related and benzodiazepine-induced effects on the states’ event-related activation profiles. Therefore, these extra analyses corroborate the robustness of our results.

Task performance

Figure 1b reports the distribution of the subjects’ mean reaction time and accuracy for the three groups (HC, MS B−, and MS B+) and for each task condition (0,1, and 2 back). The mean reaction time differs significantly between the HC and MS B− groups in the 2-back condition (one-way ANOVA test, F = 4.82, p < 0.05, multiple comparison correction via FDR, and following Tukey’s HSD Test, p < 0.05, 95% CI = [−0.107, −0.0104]), Fig. 1b. In the 1-back condition, instead, the accuracy significantly differs between groups (one-way ANOVA test, F = 10.76, p < 0.05, multiple comparison correction via FDR), Fig. 1b. In particular, the mean accuracy of the MS B+ group is significantly different from both the mean accuracy of the HC group (p < 0.05, 95% CI = [0.057, 19]) and the MS B- group (p < 0.05, 95% CI = [−0.05, −0.18]).

Fig. 1. Task description and performance measures.

a Illustration of the visual-verbal n-back task.The dark green squares highlight the target letters, and the light green squares represent the letters that are matched with the target to identify the trials. From the top, each line depicts one paradigm condition, 0, 1, and 2 back, respectively. Each letter is shown for 1 s, and the inter-stimuli time is 1.8 s. b Task Performance measures. Reaction time (RT). We plot the distributions of mean RT for the three groups, separately. For each condition, the bold line indicates the interquartile range [25–75%]. The dotted line tracks the mean for each condition and group. Mean Accuracy. The boxes indicate the interquartile range [25–75%] of accuracy across subjects for each group. The dotted line tracks the mean for each condition and group. Both for the RT and the accuracy, the statistics were performed using a one-way ANOVA test with the disease condition as a grouping variable. The black line between the two groups signs the statistical significance, *p-value < 0.05.

Description of the HMM inferred states

Spectral modes and the link with the statewise spectral content

Figure 2b illustrates the four data-driven components (spectral modes) in which the broadband frequency range [1–40] Hz was decomposed. We considered these spectral modes as the data-driven frequency bands to refer to when analysing the states’ spectral content. Spectral mode 1 covers low frequencies (approximately the theta and delta traditional bands 1–8 Hz), spectral mode 2 peaks around 10 Hz (traditionally associated with the alpha rhythm), and spectral mode 3 is broader and covers the beta and low-gamma range (15–40 Hz). Spectral mode 4 then covers mainly higher frequencies (>25 Hz). The last spectral mode is discarded because the TDE-HMM, due to the PCA step to reduce the dimensionality of the input matrix, is biased towards the low frequencies²⁶.

Fig. 2. General temporal and spectral descriptions of all the 6 inferred states.

a Temporal domain. We illustrate the steps to extract the event-related activation profiles of all the HMM states. We epoched the statewise time courses with respect to the task information, considering 200 ms before the stimulus onset and 1200 ms after the stimulus onset. Then, each trial is baseline corrected considering the [−200, −30] ms interval as baseline. Finally, we ran the generalized linear model to extract the statewise ER activations. We plotted the constant regressor for all the states, the flat lines underneath the curves represent the time window in which the correspondent state (colour-coded) is significantly different from zero (permutation test p < 0.025, number of permutations n = 1000, multiple comparison correction via maximum statistics across states and number of regressors), hence, from the baseline level. b Spectral domain. Via multitaper, we extracted the statewise power spectral density (PSD) distribution for all regions and the coherence (COH) across all pairs of regions. We plot the average PSD and COH across regions/connections per state (bold line) and the standard deviation across subjects (shaded area). Finally, we decomposed the COH into 4 components via non-negative matrix factorization (NNMF) to identify 4 data-driven frequency bands. c State-spectral mode association. Each state is associated with the spectral mode covering the frequency window where the state’s broadband PSD/COH profiles display a peak. The spectral mode represents the frequency band capturing most of the state’s spectral activity.

The spectral content of each state is associated with a spectral mode, considering the frequency window in which the state’s broadband PSD and COH profiles show a peak. In Fig. 2c, we reported these associations. The broadband PSD profile of State 1 peaks in the low frequencies, covered by spectral mode 1. States 2, 4, and 6 display a peak in their broadband PSD profiles around 10 Hz, therefore, the spectral mode associated with these states is number 2. Lastly, states 3 and 5 present a peak in the broadband COH profile around 25 Hz, therefore, their activity is linked to spectral mode 3.

Temporal and spatiospectral profiles

Following, we present the description of the HMM states in the spatial, spectral, and temporal dimensions. The latter regards the statewise event-related (ER) activation profile, reported separately for the three groups (HC, MS B−, MS B+) and the six paradigm conditions. Here, we present the results concerning the HMM inference on the whole dataset. For the spectral mode associated with the state, we report the spatial description of the state composed of the mean z-score PSD distribution over the brain and phase-coupling network.

State 1 – prefrontal theta state

The low-frequency characteristic of state 1 is localized in prefrontal regions (right and left orbitofrontal cortex, OFC, medial prefrontal cortex), the left and right anterior temporal cortices, and the posterior cingulate cortex, PCC, as shown in the PSD map, Fig. 3a. This theta prefrontal state is significantly activated early after stimulus onset, peaking ca 180 ms PST (post-stimulus time), Fig. 3b. The event-related profiles of this state differ significantly (permutation test with TFCE, p < 0.025, number of permutations n = 1000, multiple comparison correction via maximum statistics across states and group comparisons) between HC and MS B+ around the maximum peak (ca 200 ms PST) in the 0 and 1 back both target and distractor conditions; the same significant difference appears in the second part of the epochs around 600 ms PST in the 0 and 1 back distractor conditions. The same effect is observed in the description of the event-related activation profiles of state 1 for the female group, Supplementary Fig. S1, while the analysis on the male group does not report significant differences between groups, Supplementary Fig. S2. The lower ER amplitude entails a decrease in theta coupling between the recruited regions induced by the benzodiazepine treatment. A decrease in the peak amplitude, although non-significant, is also observed for the MS B- group compared to HC.

Fig. 3. Description of states 1, 2, and 3.

Each box presents the spectral content of the states in panels a (for state 1), c (for state 2), and e (for state 3). The spectral content is described by the spectral mode associated with the state, the normalized (z-score) power spectral density (PSD) map, and the brain glass with the phase-coupling network with the connections surviving thresholding via GMM. Then, for each state, we report the event-related (ER) profiles in panels b (for state 1), d (for state 2), and f (for state 3). The event-related profiles of activation of each state are reported for the three WM load conditions (0, 1, and 2 back) target and distractor, separately. In each plot, the three ER profiles are colour-coded for the three groups (HC, MS B−, MS B+) and show the mean curve (bold line) and standard deviation (colour-coded shaded areas) across subjects. The black lines underneath the curves delineate the time points in which the ER waves related to the two groups involved in the comparison (HC vs MS B−, HC vs MS B+, MS B− vs MS B+) are significantly different (permutation test with TFCE, p < 0.025, number of permutations n = 1000, multiple comparison correction via maximum statistics across states and group comparisons).

State 2 – alpha occipital activity

The alpha activity associated with state 2 is focused on the occipital cortex, and the phase-coupling networks reveal an alpha synchronization between the occipital, posterior temporal, and inferior parietal regions, Fig. 3c.

The event-related response of this state shows an early non-significant activation around 100 ms post-stimulus presentation, followed by a below-baseline reduction of activity – i.e. deactivation – between 200 and 500 ms, Fig. 3d. However, the ER activation profile of this state in the MS group does not deviate from the HC response, Fig. 3d. Therefore, the activity of this state seems to be affected neither by MS nor benzodiazepine administration.

State 3 – sensorimotor state

The beta activity associated with state 3 arises in the sensorimotor cortices, Fig. 3e. The event-related response of this state shows a negative peak around 100 ms post-stimulus, Fig. 3f. We observe that the MS group undergoing benzodiazepine treatment shows a less negative peak and an above baseline activation compared to the HC and MS B− groups; these differences result statistically significant (permutation test with TFCE, p < 0.025, number of permutations n = 1000, multiple comparison correction via maximum statistics across states and group comparisons) in the 1 back distractor condition and is qualitatively observed in the other task conditions. The same effect appears significant in the analysis of the female cohort, Supplementary Fig. S1, but not in the results regarding the male cohort, Supplementary Fig. S2.

State 4 – secondary alpha occipital state

The spectral content of state 4 was associated with spectral mode 2, therefore, alpha activity, develops in a broad occipito-parietal network, Fig. 4a. This state captures a similar activity to the one reported in state 2. However, the event-related activation profile of state 4 does not display task-evoked modulation in the whole dataset analysis, Fig. 1a, nor group differences in activation level, Fig. 4b. Therefore, we don’t consider this state as relevant to the description and understanding of the working memory dynamics.

Fig. 4. Description of states 4, 5, and 6.

Each box presents the spectral content of the states in panels a (for state 4), c (for state 5), and e (for state 6). The spectral content is described by the spectral mode associated with the state, the normalized (z-score) power spectral density (PSD) map, and the brain glass with the phase-coupling network with the connections surviving thresholding via GMM. Then, for each state, we report the event-related (ER) profiles in panels b (for state 4), d (for state 5), and f (for state 6). The event-related profiles of activation of each state are reported for the three WM load conditions (0, 1, and 2 back) target and distractor, separately. In each plot, the three ER profiles are colour-coded for the three groups (HC, MS B−, MS B+) and show the mean curve (bold line) and standard deviation (colour-coded shaded areas) across subjects. The black lines underneath the curves delineate the time points in which the ER waves related to the two groups involved in the comparison (HC vs MS B-, HC vs MS B+, MS B− vs MS B+) are significantly different (permutation test with TFCE, p < 0.025, number of permutations n = 1000, multiple comparison correction via maximum statistics across states and group comparisons).

State 5 – M300 frontoparietal state

State 5 presents higher broad spectral activity, with a peak of average COH around 25 Hz, associated with spectral mode 3, capturing the traditional beta band. This activity is distributed over a frontoparietal network including the inferior and superior dorsal PFC, the medial PFC, the left and right superior PFC, the right and left medial sensorimotor cortex (SMC), and the anterior and posterior precuneus. The synchronization of these regions via the beta rhythm results in the phase-coupling network of state 5, Fig. 4c.

State 5 task-evoked activation is significantly modulated across groups, Fig. 4d. First, an early negative deflation of the ER wave occurs around 100 ms, and this peak appears significantly (permutation test with TFCE, p < 0.025, number of permutations n = 1000, multiple comparison correction via maximum statistics across states and group comparisons) less negative for the MS B− and the MS B+ group compared to HC in the 0 back distractor conditions. The same effect arises in all task conditions when considering only the female group, Supplementary Fig. S1. Following, the event-related wave shows a positive amplitude increase between 300 and 600 ms, which we refer to as the M300 wave as it resembles the typical EEG P300 wave but is extracted in MEG data. This amplitude is significantly (permutation test with TFCE, p < 0.025, number of permutations n = 1000, multiple comparison correction via maximum statistics across states and group comparisons) higher in MS B+ compared to HC, in the time window between 400 and 600 ms PST in the 0 back target and distractor conditions, Fig. 4d. Moreover, the same group difference appears significant in all task conditions for the female cohort; in this latter analysis, this M300 activation of state 5 differs significantly between MS B- and HC in the 1, 2 target and 0, 2 distractor conditions, Supplementary Fig. S1.

State 6 – sensorimotor state

Similarly to what we explained in the parallelism between states 2 and 4, state 6 captures a sensorimotor network that is similar to state 3, Fig. 4e; however, the spectral content of this state is linked to spectral mode 2, covering alpha activity, rather than the beta activity associated with state 3. The event-related response of state 6 doesn’t show a task-induced modulation in any of the individual task conditions. However, the MS B+ group shows significantly (permutation test with TFCE, p < 0.025, number of permutations n = 1000, multiple comparison correction via maximum statistics across states and group comparisons) lower activation than HC between 400 and 600 ms PST in the 2 back target condition, Fig. 4f. The same significant effect appears in the analysis of the female cohort, not only for the MSB+ vs HC group difference but also for the MS.B− vs HC difference, Supplementary Fig. S1.

Contrasts between paradigm conditions: relevant states and group differences

The task-induced modulation of the states’ event-related activation profiles over the whole group (Supplementary Materials, Supplementary Fig. S8) indicates the role of each state in WM-specific or response-related processes. It is worth mentioning that the task-evoked activation of state 5 is significantly higher in target compared to distractor trials, therefore, revealing the role of this state in the decision-making process leading to motor response. The prefrontal activation of this state and its M300-like activation profile is also significantly modulated by WM load, which makes us link this state with high-order working memory processes such as maintenance and updating. Besides state 5, states 2, 3 and 6 are also modulated by WM load and motor response, and the related results and discussion are found in the Supplementary Materials, Supplementary Section 3.

Following, we investigated the group differences (HC vs MS B-, HC vs MS B+, and MS B− vs MS B+) within each contrast of paradigm conditions (target versus distractor, 0-2, 0-1, 1-2 target and distractor, separately). We observe no significant group differences associated with increased WM load, Fig. 5a–f. The same analysis was conducted on sex-matched groups of the dataset (only females and only males, separately), and the results are reported in the Supplementary Materials, Supplementary Section 1, Supplementary Figs. S3 and S4. Instead, the results considering the target versus distractor contrast are reported in the supplementary materials, both for the whole dataset and sex-matched cohorts, Supplementary Fig. S5.

Fig. 5. Group differences in contrasts between WM load conditions – whole dataset.

Each subplot refers to the event-related activation of a single state: a for state 1, b for state 2, c for state 3, d for state 4, e for state 5, and f for state 6. Each subplot includes six graphs presenting the event-related activation profiles for the six different paradigm conditions: 0,1, and 2 back targets and distractors, separately. Each graph also includes black lines indicating the time points where the comparisons between two groups (HC vs MS B−, HC vs MS B+, MS B− vs MS B+) are significantly different (permutation testing with threshold-free cluster enhancement, number of permutation n = 1000, p < 0.025, TFCE parameters cluster extent = 0.5 and height = 2, correction for multiple comparisons via maximum statistics across time and number of group comparisons).

Peak analysis

From the results on the group differences in the states’ event-related activation profiles, we infer that the brain networks underlying working memory functioning altered in MS and affected by the benzodiazepine administration are fundamentally two: states 1 and 5. Following, we ran the peak analysis to further study the group differences in the ER peak morphology (amplitude and latency) of these two states, considering the whole dataset, Fig. 6. These results corroborate the observations inferred from the analysis of the full states’ event-related activation profiles.

Fig. 6. Distributions of the peaks’ amplitudes and latencies for the three groups HC, MS B−, and MS B+.

Each plot presents the distributions of peak amplitude and latency for the three groups under analysis for the maximum peak of state 1 (a), and the maximum peak of state 5 (b). The significance level is drawn with a line between the two groups that differ. *p < 0.05 (permutation analysis, number of permutations = 10,000, multiple comparisons correction via FDR across a number of states, group comparisons and number of task conditions, 6).

Relationship between states’ ER significant features and behavioural data

Figure 7 reports the correlations between the behavioural data and the ER features of state 1; Fig. 8 reports the correlations between the behavioural data and the ER features of the M300 peak for state 5. For this analysis, we only considered the MS B− group to discard any benzodiazepine-induced effect and observe whether the ER features capture MS-related aspects of the brain dynamics directly linked to behavioural data. We only report the results considering the target conditions, as the results are consistent with those referring to the distractor condition, which we report in the Supplementary Materials, Supplementary Section 4, Supplementary Fig. S9.

Fig. 7. Relationship between state 1 ER peaks amplitude and latency with behavioural data for the MS B− cohort.

The behavioural data include the mean reaction time and accuracy per subject, and two neuropsychological tests, the Symbol-to-Digit modality test (SDMT) and the California Verbal Learning Test-II (CVLT-II). The scores of the neuropsychological tests are standardized (z-score). All the correlations are computed as Spearman’s correlations, and the p-values are not corrected for multiple comparisons. In each plot, the correlation coefficient (r) and the p_values (p) are displayed. Each row presents the data for a single WM load condition.

Fig. 8. Relationship between state 5 ER peaks amplitude and latency with the behavioural data for the MS B− cohort.

The behavioural data include the mean reaction time and accuracy per subject, and two neuropsychological tests, the Symbol-to-Digit modality test (SDMT) and the California Verbal Learning Test-II (CVLT-II). The scores of the neuropsychological tests are standardized (z-score). All the correlations are computed as Spearman’s correlations, and the p-values are not corrected for multiple comparisons. In each plot, the correlation coefficient (r) and the p_values (p) are displayed. Each row presents the data for a single WM load target condition.

For state 1, we observe that peak amplitudes are positively and significantly (Spearman’s correlation analysis, p-values < 0.05) correlated with the accuracy of response and negatively correlated with the reaction time in the 2 back condition. Considering the neuropsychological tests, state 1 peak amplitude is positively correlated with SDMT and CVLT-II scores; the correlation is significant only in the 2 back condition. It is worth noting the correlation between the peak amplitude and latency of state 5 and the mean accuracy of response. Interestingly, the peak amplitude is negatively correlated with the accuracy level.

Discussion

Through the time delay embedded-hidden Markov model (TDE-HMM), we decomposed the task MEG data in spectrally defined and data-driven functional brain networks^25,27. We previously implemented this technique to unveil the WM network dynamics in a healthy population²⁴. Expanding the analysis to the MS cohort, we observed that the disease itself and the ongoing benzodiazepine treatment affect the dynamics of two networks: the theta prefrontal and the M300 frontoparietal networks.

The spatial configurations of states 1 and 5 include frontal, temporal, and parietal regions, which majorly contribute to normal WM functioning⁷. FMRI studies have consistently reported an MS-induced altered activation of the same regions during WM tasks^5,10. The link between the HMM states, the WM networks, and MS findings supports the model reliability in detecting functional networks relevant to investigating WM and disease-specific processes.

However, the fMRI literature reports contradictory findings on the altered network activation^5,28. Whilst the increased activity of the prefrontal cortex in pwMS has been explained as an effect of functional reorganization to compensate for structural damage^5,8,29, a decreased frontoparietal activation and coupling were found to correlate with decreased WM task performance^9,30,31. These discrepancies might arise from the heterogeneity of the datasets, as Vacchi et al.¹⁰ demonstrated that the pattern of altered network activation strongly depends on the MS phenotype (relapsing-remitting or progressive) and disease stage^10,30. Therefore, in our research, we confined the analysis to relapsing-remitting pwMS and further split the MS cohort in pwMS treated with (MS B+) and without (MS B−) benzodiazepines, widely used symptomatic treatments in MS care to reduce anxiety, muscle spasticity, and insomnia. Although the neurophysiological alterations of benzodiazepine have previously been observed in HC^32,33, MS studies have never considered this treatment as a separate contrast, which might have also led to discrepant findings²⁰. Moreover, a recent study by Marrie et al. reported that female patients are more often administered benzodiazepine treatments than male patients³⁴. This sex bias is observed also in our dataset, as all the pwMS treated with benzodiazepine are female. For this reason and because of the sex bias within the MS population – 2:1 ratio of female:male MS patients-, we conducted the first part of the analysis (group difference analysis in states’ event-related activation profiles) for the whole dataset and, in parallel, for male and female cohorts, separately. We start by discussing the dynamics of the MS B- group, elaborating on the benzodiazepine effect separately.

State 1 depicts the early low-frequency (theta) prefrontal activity. Generated by the hippocampal-cortical circuit, theta is an excitatory rhythm that supports high-order cognitive functions, such as information integration between executive (prefrontal) and maintenance (temporal) regions, and the top-down attentional control from low to high cognitive processing^35–37. The early activation of state 1, peaking about 180 ms after stimulus onset, suggests its engagement in the first encoding stage of WM processing²⁴. We then hypothesize that state 1 performs early executive functions, crucial in (verbal) encoding and attention processes during a visual-verbal n-back task²⁴.

The decreased activation of state 1 in pwMS suggests a reduced theta coupling between the regions in pwMS compared to HC. These findings are complementary to the more traditional time-frequency analyses, reporting a decreased theta power in the hippocampus and frontal regions in pwMS^16,38. The network configuration that this work describes provides additional information about the underlying functional mechanisms impaired in MS: the decreased theta activity of state 1 indicates a decreased functional (dictated by the theta rhythm) integration between regions, consistent with what was previously reported in more static graph theory analyses^31,39.

The decoupling of state 1 could result from underlying structural damage due to atrophy and lesions, yielding functional connectivity damages. A longitudinal study reported that MS lesions develop preferentially in frontal and parietal regions, suggesting a link to the early WM impairment affecting pwMS^40,41. However, lesion development is extremely subject-dependent. Therefore, future studies should investigate the relationship between altered dynamic functional networks and lesion load and location.

Viewing the neuropsychological multi-compartment model of WM, we could link state 1 to the executive unit, given the state’s attentional control function⁶. The effect of MS on the activation of state 1 corroborates the hypothesis that the executive control unit is more strongly affected by MS than the two slave units, as previously suggested by neuropsychological research^4,8,42. Consistently, during the visual-verbal n-back task, the recruited slave unit is the phonological loop captured by state 2 (see ref. ²⁴), and we observe that its activation does not change in pwMS. This link to neuropsychology is enabled by the use of task data. Van Schependom et al. utilized the same TDE-HMM technique on resting-state MEG data and identified a frontal DMN with reduced theta power in MS compared to HC²⁰. As explained in Rossi et al., state 1 presents spatio-spectral features that resemble the frontal default mode network (DMN)^24,25. However, performing the analysis on task data allows us to link the HMM state, hence the functional network, with the specific cognitive process required during the task.

The peak analysis reveals that a correctly and rapidly performed task requires a strong activation of state 1, particularly with increasing task difficulty (from 0 to 2 back conditions). This result supports the same hypothesis for which WM performance significantly relies on the well-functioning of the executive control unit⁴². We also observe that state 1 peak amplitude is significantly correlated with California verbal learning test-II (CVLT-II) and Symbol-to-Digit modality test (SDMT). These two tests are designed to assess verbal learning and information processing speed (IPS), respectively – these processes are intrinsically akin to working memory functioning. In MS, impaired IPS was found to affect encoding, while the CVLT-II test was related to decreased attention and integrating processes^43,44. The mentioned functions were previously associated with the theta prefrontal executive activity of state 1, therefore, our results align with previous neuropsychological findings, suggesting that (working) memory impairment is associated with difficulties in attention control rather than the following retrieval stage⁴⁵.

Instead, state 5 captures a beta frontoparietal network with an ER temporal activation profile that resembles the widely used and detected EEG P300. This characteristic wave was associated with the stimuli discrimination process and represents a cognitive marker for several high-order cognitive functions, such as attention and working memory⁴⁶. The M300 state 5 is also associated with stimulus discrimination and WM-specific (updating and maintenance) processes, given that the state’s ER activation profile is modulated by WM load and motor response (Supplementary Fig. S8)²⁴.

The ER wave of state 5, the M300, does not seem to be majorly altered in the MS B− group compared to the HC. These results contradict the main findings in the MS P300 literature that repeatedly reports decreased P300 amplitudes or increased P300 latencies in pwMS^15,41,47,48.

The discrepancies that we report may arise from methodological and signal-specific issues. First, the biophysical difference between the electric and magnetic signals causes the MEG and EEG to be sensitive to different neuronal orientations, making the MEG and EEG acquisitions refer to slightly different brain sources⁴⁹. Secondly, the EEG signal suffers from volume conduction, thus, the electric signal does not homogeneously travel across the different head components (brain, meninges, skull)⁴⁹. As traditional EEG analyses are conducted at the sensor level, the EEG P300 could result from more widespread changes that sum up at the scalp level resulting in a P300 reduction in patients with MS. Instead, source-level event-related activations are more sensitive to region-specific phenomena¹⁶. Therefore, although our M300 results seem to clash with the traditional EEG P300 literature, these cannot be straightforwardly compared, as they observe the event-related brain response from different perspectives: source (MEG) versus sensor (EEG) level and network (TDE-HMM) versus single-region (traditional ER) activity.

Considering the relationship with behavioural data, our results suggest that an increased M300 latency indicates a decreased information processing speed (SDMT score), as previous EEG P300 studies have also reported⁵⁰. Therefore, the latency of the M300 state 5 seems to share the same neurophysiological meaning as the EEG P300 latency.

Instead, regarding the amplitude, the M300 shows a weak increased amplitude in MS B- group, particularly in the female group, indicating a more consistently activated beta activity in pwMS than HC. This is not the first time that an event-related analysis has reported an increased beta activity in pwMS, e.g., Kiiski et al. reported the same effect in EEG source-reconstructed data^38,51. One plausible explanation for this effect could be that the heightened arousal state in pwMS compensates for their impaired attention⁵¹. In fact, a higher arousal state is associated with increased beta activity⁵². Future studies should further investigate this beta-related effect, as well as potential sex-specific differences, to better understand this still unclear phenomenon.

Several electrophysiological (M/EEG) studies have previously observed the effect of benzodiazepines on healthy brain activity, both in event-related response and oscillatory dynamics^32,33,50. Benzodiazepines increase the GABA_A conductance of inhibitory interneurons³², which take part in the circuitry generating the theta and beta oscillatory activity^32,35,36. Considering the MS B+ group, our results show an altered ER activation profile for states 1 and 5, the networks characterized by theta and beta activity, respectively.

The ER activation of state 1 in the MS B+ group is significantly decreased compared to HC and MS B−. This alteration entails a further reduced theta power activation in state 1, caused by benzodiazepine use. Van Schependom et al. reported the same effect in resting-state MEG data²⁰. The pharmacological effect of benzodiazepine induces an increase in inhibitory currents in the GABAergic inhibitory interneurons, disrupting the excitation-inhibition balance of the cortical (medial septum-diagonal band of Broca)-hippocampal circuitry that generates the theta rhythm^35,36,53. Therefore, what we and Van Schependom et al. reported are the macroscopic network alterations reflecting the benzodiazepine effect at the neuronal circuitry level.

Instead, the beta rhythm is generated by local circuits of GABAergic inhibitory interneurons in the primary somatosensory and sensorimotor cortices⁵⁴. The increase in beta power derives from the increased GABAergic inhibition, that has systematically been reported in pharmacological, EEG, and MEG studies^32,33,53,55. Van Schependom et al. found a benzodiazepine-induced increased beta power in several TDE-HMM states²⁰. Our results show the same reoccurring effect in the two states characterized by beta activity, states 3 and 5. This last depicts frontoparietal beta activity and presents a significantly increased M300 activation in the MS B+ group compared to HC and MS B−.

Our results on the M300 clash with the EEG P300 findings in the benzodiazepine literature. Many studies report a decreased P300 amplitude as the effect of benzodiazepine^33,50,56. These discrepancies might arise from the same methodological issues that underlie the discrepant M300 findings between MS B− and HC. To the best of our knowledge, there is no understanding of the neural mechanism that generates the benzodiazepine-induced decrease in P300 amplitude. Instead, our approach enables us to link the phenomenon that generates the M300 wave and the underlying neuronal mechanism of its alteration.

Concerning the behavioural effects, benzodiazepine use was associated with reduced accuracy in performance in HC^50,55. In our results, the accuracy in the task performance is significantly altered in the MS B+ group compared to the MS B− group, Fig. 1b. Both states 1 and 5 carry out executive and WM-specific functions that can impact the task performance if altered, which becomes the case in patients with MS undergoing benzodiazepine treatment. We hypothesize that the altered benzodiazepine-induced activation observed in our results plays a role in performing the task accurately. However, due to the limited sample size (only 16 pwMS with benzodiazepine), we couldn’t further investigate the relationship between benzodiazepine use, TDE-HMM states 1 and 5 ER activation profiles, and task performance measures.

In this work, we explored the statewise event-related response that provides the first level of information on connectivity and power changes for each network. As we inferred the HMM states on the concatenated data (HC and MS), the states represent reoccurring networks appearing at the group level. This choice allowed a direct comparison between event-related responses across groups, obscuring connectivity changes between groups. If we had inferred HMM states separately for the different cohorts, we would need to run an additional step to identify which state in one inference relates to which state in the other inference. Due to the stochastic nature of the HMM inference, chances are that the states between inferences slightly differ, hindering the investigation of the disease-induced alterations. Nonetheless, in the Supplementary Materials, Supplementary Section 5, Supplementary Figs. S11–S13, we report the results for group-specific inferences – HC, MS B−, and MS B+. The results demonstrate that the TDE-HMM states supporting WM-specific processes identified in the whole dataset are reproduced and comparable across single-group inferences. Future works should evaluate a way to pursue a more detailed connectivity analysis between groups.

Here, we did not include structural data such as lesion load, atrophy measures, or information about the lesion locations. Therefore, we only speculated on the link between the altered activation of state 1 and the structural measures; instead, future studies should investigate this relationship at the group and subject levels. Further studies should also explore the variation in brain response between different benzodiazepines.

To conclude, our study examined how, during a visual-verbal WM n-back task, MS and benzodiazepine treatment affect the event-related response of spectrally defined functional networks identified through the unsupervised TDE-HMM approach. We found that MS impairs the activation of a theta prefrontal network associated with early stimulus encoding and attentional control. This result resonates with the neuropsychological literature explaining WM impairment in MS as a difficulty of the executive processes. Treatment with benzodiazepines alters the network dynamics, reflecting the treatment effect on the neuronal circuitries generating the theta and beta rhythms, respectively. Our study demonstrates that the model can identify and differentiate the disease-specific and treatment-specific effects, revealing potential new markers to assess the condition of working memory in multiple sclerosis.

Methods

Participants

The dataset includes 38 healthy controls (HC) and 70 people with multiple sclerosis (PwMS); the latter were recruited from the national MS centre (Melsbroek, Belgium) and diagnosed with relapsing-remitting MS via the revised McDonalds criteria (2010)⁵⁷. The included pwMS had an expanded disability status scale (EDSS) score equal to or below 6. To assess the effect of benzodiazepines, we split the MS cohort in two: 16 patients with (MS B+) and 54 patients without benzodiazepine treatment (MS B−), as benzodiazepines significantly alter the brain dynamics³³. The pwMS were excluded if they (a) had experienced a relapse and/or (b) were treated with corticosteroids within six weeks before the start of the study, or (c) if they carried pacemakers or dental wires or (d) suffered from epilepsy or psychiatric disorders. Table 1 presents in detail the demographics describing each dataset subgroup (based on the MEG system update, sex, and disease or treatment condition).

The patient recruitment started in 2015 and concluded in 2018. All participants signed an informed consent, and the study was approved by the ethical committees of the National MS Center Melsbroek and the University Hospital Brussels (Commissie Medische Ethiek UZ Brussel, B.U.N. 143201423263, 2015/11). All ethical regulations relevant to human research participants were followed.

Both pwMS and healthy subjects performed the BICAMS⁵⁸, a battery of neuropsychological tests including the Symbol-to-Digit modality test (SDMT) to assess information processing speed, and the California Verbal Learning-II (CVLT-II). Both tests can be relevant when evaluating working memory functioning. Table 1 reports the neuropsychological scores for the three groups.

Data acquisition

Participants underwent MRI and MEG acquisitions. The MEG machine was located at the CUB Hôpital Erasme (Brussels, Belgium) in a lightweight magnetically shielded room (MSR, MaxshieldTM, MEGIN Oy, Croton Healthcare, Helsinki, Finland). Because of a system update, 39 subjects (all female) were scanned using the Neuromag VectorViewTM system – BR1 group -, while 69 were scanned with the NeuromagTM TRIUX system (MEGIN Oy, Croton Healthcare, Helsinki, Finland) – BR 2 group. The data were acquired using 102 triplets (two planar-gradiometers and one magnetometer) of SQUID sensors. Three coils were attached to the left and right forehead and mastoid to track the head’s movements, and three sensors were used to record electrocardiography (ECG) and electrooculogram (EOG). The subject’s head shape and fiducial (nasion, left and right preauricular) points were registered using an electromagnetic tracker (Fastrak, Polhemus, Colchester, Vermont).

The MRI was acquired with a 3T Achieva scanner (Philips, Best, Netherlands) at the Universitair Ziekenhuis Brussel (Jette, Belgium). The 3D MR images were T1-weighted; the subjects were in a head-first supine position. The scan used an echo pulse sequence gradient with echo sequence TE 2.3030; the recording parameters were TR = 121 4.939 ms, flipping angle 8, field of view 230 × 230 mm², number of sagittal slices 310 with a 0.53 by 0.53 by 0.5 mm³ resolution (voxel).

The MEG (functional) and MRI (structural) data were collected within a few days (median 5, IQR 2–10 days).

Working memory task: the visual-verbal n-back

During the MEG recording, each participant performed a visual-verbal n-back task. During the n-back, a sequence of letters is displayed, and the participants must press a button (right hand) when the target letter appears. A letter identical to the n^th preceding one (n = 1,2) becomes a target. If n = 0, the letter X becomes the target. Figure 1a shows a graphic representation of the task. For each load condition (n = 0,1,2), the experiment consisted of four blocks of 20 letters (20% as target stimuli) presented pseudo-randomly. Before the actual recording, each participant underwent a training session. The letters were projected on a 6 × 6.5 cm screen located 72 cm in front of the subject. A photodiode attached to the screen precisely detected the stimulus onsets.

We discarded the missed targets and distractors with an answer. In the Supplementary Materials, Supplementary Section 6, Supplementary Table S2 and Supplementary Fig. S14 report the number of included trials and how many were discarded per paradigm condition and group (HC, MS B−, and MS B+). The recordings were cleaned from missed or wrong trials before the preprocessing, therefore, the whole data analysis, preprocessing and model inference, were conducted on the same set of trials.

The reaction time is evaluated as the time between the target onset and the moment subjects pressed the button. The accuracy of performance, instead, is measured as the ratio between the number of correctly identified trials and the total number of target trials, Fig. 1b.

MEG data preparation

MEG preprocessing

The raw data were first cleaned for background noise and head movements using the temporal extension of the maxfilter software (version 2.2 with default parameters; MEGIN Oy, Croton Healthcare, Helsinki, Finland) and then bandpass filtered in [0.1 330] Hz. The preprocessing was implemented in MATLAB 2020b using the OSL package⁵⁹. The preprocessing pipeline follows ref. ²⁷. The MEG data were coregistered to the subjects’ T1 MRI via RHINO coregistration using the subject-specific fiducial points traced by the Polhemus electromagnetic tracker. Afterwards, the data were downsampled from 1000 Hz (acquisition rate) to 250 Hz and bandpass filtered to [1,45] Hz. A notch filter at 50 Hz was introduced to remove the remaining powerline noise that could interfere with the model implemented.

Artefact removal was performed via independent components analysis (ICA), considering 62 components, and the components that correlated (r > 0.5) with the ECG or EOG recordings were discarded. Finally, the clean signals and rejected components were manually checked and visually inspected.

The data acquired by different MEG sensors (magnetometers and planar gradiometers) were normalized to overcome the difference in data variance across sensor types, following⁶⁰. Afterwards, we applied a bilateral beamformer, based on a Bayesian principal components analysis as developed in ref. ⁶⁰, to reconstruct the sources of the acquired signals. The source reconstruction was based on a single-shell forward model in the MNI space with a projection on a 5 mm dipole grid.

Parcellation

The MEG source-reconstructed data were parcelled using a data-driven 42 region-of-interest (ROIs) atlas that is reported in the Supplementary Materials, Supplementary Section 7, Supplementary Fig. S15; the same parcellation was used before in refs. ^20,24,61. This atlas covers the whole cortical surface and three subcortical regions: the precuneus, the cuneus, and the posterior cingulate cortex. The first principal component between the voxels’ time series under a single ROI was assigned as the time course of the ROI itself. The ROIs’ time courses were orthogonalized by multivariate symmetric leakage correction to discard signal leakage across parcels⁶². Lastly, we applied the sign flipping algorithm, as proposed by Vidaurre et al. ⁶¹, to overcome the sign ambiguity issue affecting beamformed data.

TDE-HMM

We implemented the time delay embedded-hidden Markov model (TDE-HMM) to extract dynamic functional networks from the MEG preprocessed data. For a thorough mathematical description of the model, we refer to refs. ^25,63. We have previously implemented the same technique to unveil the network dynamics underlying working memory processing in healthy subjects²⁴, so we only describe the method conceptually in what follows.

Time delay embedded - Hidden Markov model

Generally, a hidden Markov model describes the empirical data (MEG) as alternating activations of a discrete number of hidden states, the functional brain networks, Figure S16 in the supplementary materials. As a Markovian model, the activation of a state at time t depends directly on the observed data at time t and the adjacent state activated at time t-1⁶⁴. Here, we use the time delay embedded-HMM, in which each state is modelled by a (low-rank) Gaussian distribution across regions and time points, spanning through a time window of 60 ms (15 time points), thereby capturing specific spectral patterns across regions. The Gaussian distribution is assumed to have zero mean so that the covariance across regions and time points is the only state parameter. The estimation of the model parameters is done using stochastic Bayesian variational inference^61,63.

We ran the TDE-HMM at the group level, i.e. on both the HC and MS concatenated data, so that the brain states are straightforwardly comparable between groups. Although the states are inferred at the group level, the statewise time, space, and frequency profiles can be evaluated per subject (as explained below), which allows us to compare the states’ features between groups. This is analogous to the dual estimation process⁶⁵. Regarding the number of states to infer, we set apriori the number to 6, based on previous work^1,21,23,24. Importantly, the brain states are inferred without entering any information on the task. The model’s output that we consider is the posterior probability (probability of activation) of each state. This gives us the probability of a state being active at each point of time – forming the states’ time series.

Temporal description of the event-related activation profiles

We epoched the statewise time courses (probability of activation per time point) with respect to the stimulus onset, creating epochs of 1400 ms, [-200, 1200] ms. Each epoch was baseline corrected, considering as baseline the [-200, -30] ms time window before the stimulus. From these epoched data, we ran two analyses. First, a whole dataset (no disease- or treatment-related group distinction) analysis to investigate the general states’ task-evoked activation profiles, revealing which states are significantly modulated by the stimulus presentation, increasing WM load, and motor response. This investigation was conducted via a generalized linear model (GLM) with 8 regressors: a constant regressor, three paradigm conditions (0, 1, and 2 back target), a regressor describing the contrast between target and distractor trials, and three regressors including the contrasts between different working memory load conditions (1-0, 2-0, 2-1). Plotting the constant regressor, Fig. 2a, we identify the task-relevant states as those that significantly modulate their activation (time course) in response to the task²⁴. Moreover, we observe which states are associated with WM load and motor response (supplementary materials, Figure S8). The second analysis conducted on the epoched data concerns the group differences (HC vs MS B- vs MS B + ) in the statewise activation profiles across different task conditions, which is explained in the following section on statistical analysis.

States’ spatio-spectral description

We extracted the spectral description of each state and subject separately. First, the MEG data were multiplied by the statewise posterior probabilities – resulting in the statewise MEG time courses. From this, a non-parametric multitaper was used to compute the power spectral density (PSD) for each brain region and the coherence (COH) between each pair of brain regions. As the multitaper is run singularly for each subject, we obtained the subject and state-specific spatial (PSD) and phase-coupling (COH) maps, Fig. 2b^25,63.

The PSD and coherence were first extracted in the broad frequency band [1, 40] Hz. Afterwards, we decomposed the frequency band into four components (spectral mode) in a data-driven way using a non-negative matrix factorization of the coherence measure, as described in^25,27. The number of components (4) was chosen considering the number of conventional frequency bands included in the 1–40 Hz spectrum: theta/delta [1–8] Hz, alpha [8–12] Hz, beta [12–25] Hz, and low gamma [25–40] Hz (ref. Figure 2b).

By multiplying each spectral mode for the PSD and COH of each subject and state, we extracted the statewise spatial (PSD) and phase-coupling (COH) maps for each data-driven spectral mode. Then, per state, we identify the frequency window where the broadband PSD/COH peaks appear, and we associate the spectral mode covering that frequency window with the state under consideration. This step is illustrated in Fig. 2c.

In Figs. 3 and 4, we only report the spatial activation and connectivity (coherence) maps for the spectral mode identified as the one capturing at best the state’s spectral content. Solely with the purpose of visualization, we z-normalized the power spectral density (PSD) maps of the whole group. No thresholding was applied here, but we maintained the PSD distribution over the whole brain, providing a complete description of the state spatial activation.

The coherence values plotted in Figs. 3 and 4 result from the product of the broadband coherence values reported in Fig. 2b (values that fall in the 0-1 range) with the weights characterizing the spectral mode distribution chosen to represent the state spectral content. In this way, the coherence connections plotted in Figs. 3 and 4 are associated with the considered spectral mode and those that were associated with other frequencies are suppressed. Intending to plot a clear network configuration, the coherence connections were thresholded individually per state and spectral mode, utilising a Gaussian mixture model (GMM). This technique models the population of connections per state and spectral mode as a mixture of two Gaussian distributions. The first distribution covers the most probable mean of connectivity strength, and the second one considers those connections with strength higher than 95% of the overall mean strength. Only the connections that fall within the second distribution are plotted. This step has the only goal of illustrating a clear coherence map of the state at the group level, so no further statistics is conducted.

Statistical analysis

Group differences in states’ ER activation profiles

To investigate the group differences (HC vs MS B- vs MS B + ) in the event-related activations of the HMM states, we first took the average task-evoked response per subject and paradigm condition, considering the 0, 1, and 2 back, target and distractor (6 conditions), separately. The contrasts between paradigm conditions were also evaluated: target versus distractors (all WM loads together), 0-2, 0-1, and 1-2 targets and distractors, separately. We implemented non-parametric permutation testing (number of permutations = 1000, p-values < 0.025) with a threshold-free cluster enhancement transformation (parameters cluster extent e = 0.5, and cluster height h = 2 following⁶⁶) to evaluate the difference of the states’ temporal activation profile (occupancy level) between the three groups (HC, MS B-, and MS B + ). The permutation test was computed per state, per paradigm condition and contrast, and per time point over the time window [0, 1000] ms. Then, we corrected the tests for multiple comparisons via maximum statistics (maximum over time and number of group comparisons). This analysis provided the time points within an epoch, in which, for each state singularly, the event-related activations of the three groups (HC, MS B-, and MS B + ) significantly differ.

ER features – amplitude peak and latency

We ran a peak analysis to further explore the morphology of the states’ event-related activations and the differences between groups. We extracted the maximum peaks of states 1 and 5 as these displayed the most significant differences between groups in the ER activations. For state 1, the maximum positive peak was extracted in the [50, 400] ms time window, whilst the M300 maximum positive peak of state 5 was extracted in the [300, 800] ms time window. We are interested in observing the amplitude and latency for each peak.

We ran permutation testing (number of permutations = 10000, multiple comparison correction via FDR) to reveal the statistical significance of the difference in mean amplitude and latency of states’ peaks between groups (HC, MS B-, and MS B + ) across paradigm conditions (0, 1, 2 back targets and distractors). This peak analysis served to link the findings over the states’ altered ER activations in the MS groups to the observable behavioural data. After extracting the peaks’ amplitude and latencies, we correlated them (Spearman correlation) with the behavioural data: mean reaction time, mean accuracy, SDMT, and CVLT-II tests. The correlation was run per state, peak, and task condition (0,1, 2 back targets), separately. Given the exploratory nature of this analysis, we did not correct the correlations for multiple comparisons. Nonetheless, the total number of correlations performed is 48, and we provide a better overview of which correlations are explored: 2 states x 1 peak per state x 2 morphology parameters (latency and amplitude) per peak x 3 paradigm conditions (0,1, and 2 back) x 4 behavioural variables.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

C.R. conducted the analysis and wrote the manuscript. J.V.S. was the main supervisor of the work and helped write and review the manuscript. D.V., L.C., and G.N. gave inputs for the analysis and provided feedback in the writing process. M.W., M.B.D., M.D and F.A. provided comments on the work. All authors approved the submitted version.

Peer review

Peer review information

Communications Biology thanks Jan Kujala and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editor: Benjamin Bessieres. A peer review file is available.

Data availability

The MEG and MRI data for this study are not publicly available due to privacy restrictions following the General Data Protection Regulation (EU) 2016/679. Researchers interested in collaborating on these data are welcome to contact the senior authors (Prof. Jeroen Van Schependom and Prof. Guy Nagels). The source data to reproduce the figures in this paper are included in the Supplementary Data 1–3.

Code availability

The analyses were conducted in MATLAB, utilizing the freely accessible HMM-MAR package which can be found here: https://github.com/OHBA-analysis/HMM-MAR.This package belongs to the OSL (OHBA Software Library) toolbox that can be consulted here²⁷. In particular, the analysis we implemented was based on the work presented by ref. ²⁷. The scripts containing the full pipeline (MEG preprocessing, HMM inference, and data analysis with GLM and spectral decomposition), and that can be used to reproduce the analysis conducted in this work, can be found here: https://github.com/OHBA-analysis/Quinn2018_TaskHMM. For more details on the analysis scripts, contact the corresponding author (Chiara Rossi).

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s42003-024-07283-2.