Coupling between neural oscillations and white matter integrity reveals cognitive computational profiles following COVID-19

Abstract

Persistent cognitive symptoms following SARS-CoV-2 infection have been widely reported, yet their underlying neural mechanisms remain poorly understood. Here, we applied an integrative framework—combining EEG, diffusion MRI, and computational cognitive modeling—to investigate brain–behavior relationships in patients recovered from COVID-19 (n = 70) and other non-COVID respiratory infections (n = 26). Participants completed tasks probing cognitive control, working memory, and decision-making. Across tasks, individuals with prior COVID-19 showed altered low-frequency oscillatory activity and performance deficits, despite mild or moderate acute illness. These oscillatory alterations were linked to reduced white matter integrity in distinct long-range tracts, including the cingulum and thalamo-occipital fasciculi, depending on the cognitive computation involved. Crucially, model-based clustering revealed two distinct neurophysiological profiles among post-COVID participants, reflecting different combinations of structural and functional alterations. These profiles were not explained by clinical severity markers or other symptoms like anosmia, suggesting the emergence of neurocognitive phenotypes beyond binary classifications. Our findings identify a generalizable principle whereby specific brain network disruptions underlie cognitive dysfunction, highlighting biomarkers that may guide personalized intervention strategies. Beyond COVID-19, this multimodal framework offers a scalable approach for uncovering structure–function–computation coupling in other post-infectious or neuroinflammatory conditions, with broad relevance for the prevention and treatment of cognitive impairment.

Introduction

The COVID‑19 pandemic, caused by SARS‑CoV‑2, has infected over 777 million people and claimed more than 7 million lives¹ worldwide, and while acute symptoms and respiratory complications have been extensively characterized, the long‑term neurological consequences, such as potential exacerbation of neurodegeneration and cognitive loss, remain poorly defined^2,3. Patients recovered from COVID-19 frequently present neurological and cognitive sequelae, even in individuals with initially mild respiratory symptoms⁴. There is no clear, uniform pattern of cognitive impairment following COVID‑19, but numerous studies report deficits in executive function, decision‑making, attention, and memory^5,6. A range of risk factors appear to influence post‑COVID neurological and cognitive outcomes, including older age, preexisting medical conditions, severity of the initial infection, immune response, and early markers of nervous system involvement such as anosmia^7,8,9.

Regarding the mechanisms of SARS‑CoV‑2 infection, it has been reported that the virus can infect neurons via receptors like NRP1 and ACE2, triggering innate immune responses, altering amyloidogenic processing, and up‑regulating amyloid‑β through γ‑secretase modulation^10,11. The presence of SARS‑CoV‑2 in postmortem human retina tissue and olfactory mucosa underscores its ability to reach the nervous system, raising critical questions about potential cognitive deficits and neurodegenerative risk following infection^10,12,13.

Neuroimaging studies in humans have suggested that cognitive deficits following COVID-19 stem from disruptions in neural structure, particularly white matter integrity^{8,14,15,16,17,18,19}. Specifically, studies involving diffusion MRI reveal that even mild COVID‑19 can result in lower fractional anisotropy (FA) and higher mean diffusivity (MD) across widespread white matter tracts—effects that can persist up to one year post-infection^{8,19,20,21,22}. These structural alterations may underlie the observed functional abnormalities in brain oscillatory activity. In healthy individuals, alpha oscillations play a causal role in modulating spatial attention and visual processing^23,24 by implementing a mechanism of attentional suppression that increases over cortical regions responsible for processing distracting information^25,26,27. Similarly, delta/theta oscillations serve as default communication patterns between prefrontal and visual cortices in healthy subjects, mediating distinct functional roles in spatial and nonspatial top-down attention and working memory control^28,29. Importantly, COVID-19 patients exhibit disruptions in these oscillatory patterns, showing slowing in resting-state EEG signals with reduced alpha rhythms and increased slow-wave activity (delta and theta), similar to patterns observed in early neurodegenerative diseases during rest^30,31. These oscillatory abnormalities have been associated with altered functional connectivity and structural changes in regions such as the hippocampus^32,33.

Emerging evidence underscores that SARS‑CoV‑2 infection can cause detectable and potentially long-lasting damage to the central nervous system, with significant public health implications as the population of COVID‑19 survivors increases and additional risk factors for neurodegeneration become more prevalent. Such effects have been evidenced even among adolescents and young adults, who report persistent symptoms including fatigue, headache, impaired concentration, and memory deficits^16,32,34. These findings signal an urgent need for long-term neurological monitoring and early interventions to mitigate potential waves of neurodegenerative disease emerging in the aftermath of the pandemic. In this context, identifying mechanisms that link structural damage to functional cognitive consequences is a crucial step in developing scientifically grounded interventions. A key challenge in this endeavor lies in the heterogeneous and spatially distributed nature of neural injury—ranging from focal lesions to widespread disruptions—which complicates the consistent identification of vulnerable brain networks, as well as the characterization of their associated cognitive manifestations. This variability underscores the critical need to incorporate multimodal approaches, ranging from advanced neuroimaging to comprehensive cognitive assessments, to guide evidence-based therapeutic strategies. Moreover, recent advances in cognitive computational modeling, along with increasingly sensitive methods for detecting structural and functional brain alterations, have enabled the identification of specific computational disruptions in oscillatory activity and network dysfunctions that underlie behavioral deficits in patients^35,36,37,38.

In this study, we investigated oscillatory brain activity patterns underlying specific cognitive computations, focusing on alterations observed in patients recovered from COVID-19. We then examined how these oscillatory features relate to the microstructural integrity of brain networks, aiming to identify specific mechanisms that underlie such alterations and could serve as potential targets for individualized interventions. Finally, we explored whether patients could be clustered based on distinct patterns of neural damage, as a step toward designing efficient therapeutic interventions personalized to optimize cognitive performance.

Our results reveal widespread alterations in cognitive computations among individuals recovered from COVID-19. These deficits in cognitive processing are linked to altered oscillatory dynamics, particularly in low-frequency bands, depending on the nature of the computation involved. Each oscillatory feature associated with a specific cognitive process was matched to distinct microstructural disruptions in white matter tracts. Finally, cluster analysis distinguished two main profiles among post-COVID-19 patients, characterized by different oscillatory patterns that, in turn, reflect differences in cognitive computations and white matter tract integrity, pointing to precise mechanisms that could be targeted through individualized therapeutic interventions^39,40.

Results

Participants

This study included participants from a prior MRI study⁸ invited to a follow-up EEG session involving three cognitive tasks: a behavioral inhibition and conflict expectation task (Go/Nogo, GNG) task, a working memory (WM) task, and a reversal learning (RL) task. The cohort consisted of individuals with mild to moderate respiratory symptoms (according to clinical severity criteria), who did not have respiratory failure or the need for invasive ventilation, and had remained asymptomatic for at least four weeks prior to assessment. Participants were 96 patients (aged 19–66 years; median = 39) including 70 COVID-19 recovered cases (confirmed by SARS-CoV-2 PCR) and 26 non-COVID-19 patients (with a non-COVID acute respiratory disease confirmed by at least two negative PCR tests and no history of COVID-19). We evaluated cognitive, structural, and functional alterations and their association with clinical diagnosis and profiles. We focused on two clinical factors: anosmia during the acute phase of the infection (An, including hyposmia and microsmia) as a neurological marker, and hospitalization during the acute phase of the infection (HR) as an indicator of the severity of respiratory symptoms. These factors were used as grouping variables in a multiple linear regression analysis. There were no significant age differences between post-COVID-19 patients and controls, nor between post-COVID-19 patients with or without anosmia or hospitalization (|t|< 0.7, p > 0.5, |r|< 0.1), except for those who were hospitalized during the acute phase (t = 5.4, p = 1 × 10,r ≈ 0.5). Enrollment occurred 2–27 months post diagnosis (median = 9 months), with no group difference between groups (|t|< 1.7, p > 0.06, |r|< 0.17) except for patients that were hospitalized during acute episodes (t = 6.4, p = 5 × 10^–10_, r ≈ 0.57). Given those differences, all results were correct using a model that includes age and time from diagnosis as a control variable. Other variables, such as education and sex, did not differ between groups (βs < 0.2, ps > 0.3, |r|< 0.1).

Behavioral results of computational cognitive processing

For all tasks, we computed behavioral indices and latent variables derived from cognitive computational modeling (as described in prior work^8,35,41) to compare participants based on their clinical factors (see previous section).

In the Go/No-Go task, we assessed 1) detectability (d’) of Go stimuli as a general measure of behavioral inhibition (see Methods), and 2) expectation of conflict as both a behavioral index⁴² and a latent variable explaining the computational processing of proactive cognitive control³⁵ (see Methods). Participants demonstrated accurate task performance (91.7% of correct responses for GO stimuli and 80.6% of correct inhibition for Nogo stimuli, mean d’ = 2.8, 95% CI [2.6, 3.0], p < 2 × 10^–16), with no significant between-group differences (GLM: all |t|< 0.98, all p > 0.3, |r|< 0.3 ). In the non-COVID-19 group, we observed significant reaction time slowing across consecutive Go trials, with mean RT increasing by 15.8 ms (95% CI [10, 21]; Wilcoxon V = 859, p = 1.3 × 10^–17, r = 0.54) when comparing the first two Go stimuli of a sequence to subsequent Go stimuli. post-COVID-19 patients exhibited attenuated slowing when compared to the control group (β = −0.34, t = -2.1, r ≈ −0.24, p = 0.03), indicating reduced proactive cognitive control in the post-COVID-19 patient group. No other regressors significantly modulated the sequential slowing effect (all |t|< 0.84, all p > 0.4, |r|< 0.07). We then computed a cognitive model of conflict expectation (Ex), reflecting conflict probability learning and sequence building, revealing consistent effects. A generalized linear mixed model (GLMM, log-normal distribution of reaction-time) showed significant main effects of Ex (β = 0.093, t = 3.1, r ≈ 0.2, p = 0.001) and a significant Ex × COVID-19 interaction (β = −0.053, t = -2.6, r ≈ 0.19, p = 0.007). Bayesian estimation using a drift diffusion model confirmed these findings, with Ex showing a median effect of 14.6 (HDI [0.3, 27.7], p_MCMC = 0.04) and the Ex × COVID-19 interaction yielding a median of −10.1 (HDI [−18.9, −1.3], p_MCMC = 0.02).

In the WM task, we analyzed performance using logistic mixed-effects models. As expected, accuracy declined with increasing memory load (β = -0.63, SE = 0.04, t = -14.5, r ≈ -0.53, p < 1 × 10⁻¹⁶). This load-dependent impairment was significantly more pronounced in post-COVID-19 individuals (interaction: β = −0.20, SE = 0.06, t = −3.1, r ≈ −0.20, p = 0.001) than in those from the non-COVID-19 group. No other clinical variables significantly modulated the effect of memory load (all |t|< 0.6, p > 0.5, |r|< 0.1). Bayesian estimation yielded converging results: the posterior distribution of the interaction between memory load and COVID-19 diagnosis had a median of -0.17, a 95% HDI of [−0.31, −0.06], and a p_MCMC = 0.004.

Finally, in the RL task, we partially replicated the findings previously reported in the earlier study that included the same RL task solved by the participants during the fMRI session⁸. We employed a decision-making model grounded in prospect theory and estimated the probability of winning for each deck using a Rescorla-Wagner learning algorithm⁸. From each run, we derived separate learning rates for positive and negative feedback (see Methods). Following the previous analysis frameworks, we focused on the learning rate derived from negative feedback. Using a general linear model (GLM) to assess group differences while controlling for relevant clinical variables, we found that post-COVID-19 patients exhibited a significantly lower learning rate compared to non-COVID-19 patients (β = −0.39, SE = 0.18, t = −2.1, r ≈ −0.57, p = 0.03) but post-COVID-19 patients with anosmia presented a higher rate that post-COVID-19 patients without anosmia, indicating impulsive change when facing a negative feedback (β = 0.41, SE = 0.16, t = 2.4, r ≈ 0.42, p = 0.01). Bayesian estimation confirmed the modulation of this parameter in post-COVID-19 patients with anosmia (posterior median = 0.27, 95% HDI = [0.01, 0.55], p_MCMC = 0.04), but not for COVID-19 diagnosis per se (posterior median = −0.16, 95% HDI = [−0.48, 0.12], p_MCMC = 0.28).

Electroencephalography results related to computational cognitive processing

We conducted two primary statistical analyses on EEG data to investigate cognitive dynamics in post-COVID-19 and non-COVID-19 control patients. First, based on prior evidence linking brain alterations to COVID-19 anosmia, we categorized participants into three groups: (1) non-COVID-19 controls, (2) post-COVID-19 patients who never experienced anosmia, and (3) post-COVID-19 patients who experienced anosmia during the acute phase of the disease. This grouping approach was used primarily for descriptive purposes, allowing a straightforward inspection of the main oscillatory patterns associated with anosmia. However, it did not account for potential confounding variables. Therefore, the confounder-adjusted general linear model (GLM) was established as the main statistical analysis. This model was performed at the group level using the same main contrasts—COVID-19 (present/absent) and anosmia (present/absent)—while statistically controlling for potential confounders, including age, hospitalization during the acute COVID-19 infection due to respiratory symptoms, and the interval between COVID-19 diagnosis and EEG assessment. This approach ensured that the EEG results reflected effects specifically attributable to COVID-19-related factors rather than demographic or clinical variability. This second analysis mirrored the structure used in the behavioral analysis and provided the primary inferential results reported in the EEG section.

Expectation of conflict

For the GNG task, we performed a model-based analysis of neural power modulation (see methods and prior work^{43,44,45,46,47}) based on behavioral results. Specifically, we operationalized proactive cognitive control by modeling each participant’s conflict expectation (see methods), then regressed the estimated latent variable against spectral power changes to isolate relevant neural activity following Go stimuli. Focusing on a fronto-central region of interest (ROI: E7, E4, E53), we identified task-related oscillatory dynamics consistent with prior work^35,42, as depicted in Fig. 1 through the first step of the two-tiered highlighting scheme in the time–frequency charts. Specifically, we found a post-stimulus theta (3–6 Hz) enhancement following Go stimuli, and pre- and post-stimulus beta (13–30 Hz) around the Go stimulus (Fig. 1, time–frequency plots). Following behavioral findings, nonparametric Kruskal–Wallis (KW) and cluster-based permutation (CBP) tests revealed significant between-group differences in theta power (the second step of the two-tiered highlighting scheme in Fig. 1). These effects remained significant after controlling for confounding factors in the group-level GLM analysis (COVID regressor, CBP-corrected p = 0.002, mean adjusted Cohen’s d = 0.56, mean partial r = 0.27). Topographic distribution of this analysis further demonstrated a progressive attenuation of theta modulation across groups: strongest in controls, intermediate in post-COVID-19 patients without anosmia, and weakest in those post-COVID-19 patients with anosmia. This gradient suggests a dose-dependent relationship between COVID-19-related olfactory dysfunction and disruptions in cognitive control activity.

Fig. 1

EEG Results. (A) Oscillatory power from frontal electrodes correlates with conflict expectation, as estimated from a behavioral model. The main task effect is shown and statistically highlighted using a two-tiered approach: the first layer indicates regions significantly modulated by conflict expectation (Kruskal–Wallis test followed by Cluster-Based Permutation testing; cluster-corrected p < 0.001). The second layer indicates regions showing significant group differences in this modulation (cluster-corrected p = 0.001). (B) Topographic maps (topoplots) illustrate the distribution of theta power correlated with conflict expectation for each patient group. (C) Oscillatory power from frontal electrodes correlates with memory load. The main task effect is shown and statistically highlighted using a two-tiered approach: first layer, regions significantly modulated by memory load are marked (Kruskal–Wallis test followed by cluster-based permutation testing; cluster-corrected p < 0.001). The second layer indicates regions showing significant group differences in this modulation (cluster-corrected p < 0.001). (D) Topoplots illustrate the distribution of theta power correlated with memory for each patient group. (E) Oscillatory power from frontal electrodes correlates with perceived uncertainty, as estimated from a behavioral model. The main task effect is shown and statistically highlighted using a two-tiered approach: the first layer indicates regions significantly modulated by uncertainty (Kruskal–Wallis test followed by cluster-based permutation testing; cluster-corrected p < 0.001). The second layer indicates regions showing significant group differences in this modulation (cluster-corrected p = 0.001). (F) Topographic maps (topoplots) illustrate the distribution of theta power correlated with uncertainty for each patient group. (A–C) Small black-to-red topoplots indicate statistical comparisons at the electrode level between groups, using the Kruskal–Wallis test (on the right) or specific pairwise contrasts derived from a GLM controlling for hospitalization, age, and time since diagnosis in all participants (bottom; see Methods).

Memory load

For the WM performance, following behavioral results, we conducted a model-based analysis of neural power modulation using memory load as the regressor for spectral power (following Figueroa-Vargas et al⁴¹). Consistent with behavioral findings, the control group exhibited significant theta/alpha-band modulation related to memory load (5–10 Hz) during the maintenance period in left frontal electrodes (ROI: E11, E13, E12, E9). Notably, post-COVID-19 patients with anosmia showed significantly reduced theta modulation compared to other groups (cluster corrected p < 0.001, KW and CBP tests), suggesting impaired maintenance, reflected as phase neural synchronization (Fig. 1). These effects remained significant after controlling for confounding factors in the group-level GLM analysis (ANOSMIA regressor, CBP-corrected p < 0.001, mean adjusted Cohen’s d = 0.70, mean partial r = 0.33). Group comparison using this GLM shows a small but significant positive difference between the control group and post-COVID-19 patients without anosmia. Interestingly, post-COVID-19 patients with anosmia show a significant decrease in theta activity compared with both control and post-COVID-19 patients without anosmia. These results revealed a non-linear relationship in oscillatory modulation, showing the greatest impairment in post-COVID-19 patients with a history of anosmia.

Decision-making variables

In the RL task, we conducted a model-based analysis of neural power modulation, focusing on two key decision variables influenced by the learning rate, which were found to differ between patient groups in the behavioral results. Specifically, we evaluated the value of the selected option and the perceived uncertainty associated with the current decision (see Methods). While task-related modulation was observed for the value variable, no significant differences emerged between groups. In contrast, analysis of the uncertainty variable revealed a reduction in power over parietal and frontal electrodes following option presentation. This finding aligns with the proposed causal role of the parietal and frontal cortex in processing ambiguity and uncertainty³⁷. Similar to the previous task, these effects remained significant after controlling for confounding factors in the group-level GLM analysis (COVID regressor, CBP-corrected p = 0.007, mean adjusted Cohen’s d = 0.57, mean partial r = 0.27), with differences observed among the three patient groups.

White matter integrity and brain oscillations associated with cognitive computations

The oscillatory results suggest that COVID-19 impacts cognitive computations through alterations in brain oscillatory activity. However, these changes are not consistently associated with clinical factors such as anosmia across the different cognitive tasks. This finding points to patient-specific patterns of neural alterations that extend beyond the clinical features previously assessed. We conducted the following analyses to determine whether these individual and specific alterations in cognitive computation are also reflected in distinct structural brain changes. Building on the preceding results and previous findings suggesting that white matter integrity is particularly sensitive for detecting brain injury caused by COVID-19, we investigated whether specific oscillatory profiles are associated with distinct patterns of white matter alterations.

Specifically, we first identified the oscillatory brain activity patterns most prominently across the entire sample and related to altered computational cognitive processes in post-COVID-19 patients. Next, we extracted fractional anisotropy (FA)—a measure of white matter integrity derived from diffusion-weighted imaging (DWI; see Methods)—from 36 long-range white matter tracts. We then applied independent component analysis (ICA) to reduce dimensionality and identified the components explaining 90% of the variance in the sample. Finally, we used a Bayesian Lasso approach to identify which components of white matter integrity were most strongly associated with the oscillatory activity linked to specific cognitive deficits. Clinical variables such as diagnosis, anosmia, and hospitalization were also included in the model to better characterize the interaction between oscillatory dynamics and white matter alterations that may underlie the cognitive impairments observed, providing a more nuanced view than categorical group comparisons alone.

Specifically, we first identified the oscillatory brain activity patterns most prominently expressed across the entire sample and related to altered computational cognitive processes in post-COVID-19 patients. Next, we extracted fractional anisotropy (FA)—a measure of white matter integrity derived from diffusion-weighted imaging (DWI; see Methods)—from 36 long-range white matter tracts. We then applied independent component analysis (ICA) to reduce dimensionality and identified the components explaining 90% of the variance in the sample. This dimensionality reduction step effectively limits the number of independent tests and mitigates the risk of inflated false positives arising from correlated tract-level measures. Finally, we used a Bayesian LASSO approach to identify which components of white matter integrity were most strongly associated with the oscillatory activity linked to specific cognitive deficits. The Bayesian LASSO imposes an L₁ penalty that shrinks non-informative coefficients toward zero and performs automatic feature selection, thereby providing a model-based correction for multiple comparisons within a single inferential framework. Clinical variables such as diagnosis, anosmia, and hospitalization were also included in the model to better characterize the interaction between oscillatory dynamics and white matter alterations that may underlie the cognitive impairments observed, providing a more nuanced view than categorical group comparisons alone. Together, this ICA–Bayesian LASSO pipeline ensures that significant multimodal associations reflect robust, reproducible relationships rather than chance findings driven by the large number of neural and clinical features examined.

For expectation-related computations reflected in theta oscillations when analyzing the GNG task, we found that the strongest modulation across the entire sample mirrored the group differences observed in the previous analysis (Theta: [3–6] Hz, [0.6 0.6]s, see Fig. 2A). We then used this oscillatory signal, recorded from fronto-central electrodes, to examine its association with white matter integrity. Independent Component (IC) #6 showed a significant association with theta oscillatory activity (median = 0.21, 95% HDI [0.03, 0.39], p_{LASSO_MCMC} = 0.018). Using the components with the strongest influence in the Bayesian Lasso regression (see Methods, ICs #6, #11 p_{LASSO_MCMC} = 0.074, and #18 p_{LASSO_MCMC} = 0.083), we identified the white matter tracts contributing most to this association. Notably, the right thalamus-parietal fasciculus emerged as the tract most strongly linked to theta activity (median = 0.095, 95% HDI [0.014, 0.17], p_{LASSO_MCMC} = 0.0239, Fig. 2B, C).

Fig. 2

Coupling between cognitive computations, oscillatory activity, and white matter integrity. (A) Brain source of theta-band activity ([3–6] Hz, [0.6–0.8] s) related to conflict expectation (see Fig. 1a). The source is thresholded at q < 0.01 (FDR-corrected) for visualization. (B) Thalamocortical white matter fasciculi. (C) Association between conflict-expectation-related theta activity and white matter integrity (fractional anisotropy, FA) in the thalamocortical fasciculus, assessed using Bayesian SLL regression. (D) Brain source of theta-band activity ([2–5] Hz, [2.3–2.8] s) related to memory load during the maintenance phase (see Fig. 1b). Source thresholded at q < 0.01 (FDR-corrected). (E) Left and right cingulate fasciculi (I and II). (F) Association between memory-load-related theta activity and FA in the left cingulate fasciculus, assessed using Bayesian SLL regression. (G) Brain source of theta-band activity ([4–7] Hz, [0–0.2] s) related to uncertainty during decision-making (see Fig. 1b). Source thresholded at p < 0.05 (uncorrected). (H) Uncinate and thalamo-occipital fasciculi. I Association between uncertainty-related alpha activity and FA in the right thalamo-occipital fasciculus, assessed using Bayesian SLL regression.

Following the same procedure, we identified oscillatory modulation related to memory load computations at theta and delta frequencies ([2–5]Hz, [2.3 2.8]s, see Fig. 2D) when analyzing the WM task, consistent with the lower frequency range where group differences were observed. Interestingly, the same independent component (IC #6) emerged as significant, but with the opposite sign (median = −0.20, 95% HDI [−0.39, −0.02], p_{LASSO_MCMC} = 0.0217). Using the components with the strongest influence in the Bayesian Lasso regression (see Methods; ICs #6, #9 p_{LASSO_MCMC} = 0.050, and #14 p_{LASSO_MCMC} = 0.066]), we identified the cingulum fasciculus as the tract most strongly associated with this modulation. Specifically, the following subcomponents of the cingulum bundle showed significant associations: the left cingulum cingulate fasciculus (median = 0.087, 95% HDI [0.01, 0.16], p_{LASSO_MCMC} = 0.019, see Fig. 2E, F), the left cingulum cingulate fasciculus 2 (median = 0.090, 95% HDI [0.006, 0.18], p_{LASSO_MCMC} = 0.035), the right cingulum cingulate fasciculus (median = 0.070, 95% HDI [0.01, 0.14], p_{LASSO_MCMC} = 0.044), and the right cingulum cingulate fasciculus 2 (median = 0.090, 95% HDI [0.005, 0.18], p_{LASSO_MCMC} = 0.041).

For learning under uncertainty, we identified theta oscillations across the entire sample when analyzing the RL task, consistent with the results presented in Fig. 1 ([4–7]Hz, [0 0.2]s, Fig. 2G), and localized to a frontal region of interest (ROI). Independent Component (IC) #4 was significantly associated with this modulation (median = −0.32, 95% HDI [−0.50, −0.15], p_{LASSO_MCMC} = 5 × 10⁻⁴). The main ICs contributing to the Bayesian Lasso regression (ICs #4, #1 [p_{LASSO_MCMC} = 0.14], #15 [p_{LASSO_MCMC} = 0.052], #18 [p_{SSL_MCMC} = 0.060], and 190.15 [p_{LASSO_MCMC} = 0.15]) revealed that the tracts most strongly associated with this modulation were the thalamus–occipital fasciculus (median = –0.18, 95% HDI [–0.29, –0.07], p_{LASSO_MCMC} = 4 × 10⁻⁴, Fig. 2H, I) and the left uncinate fasciculus (median = –0.09, 95% HDI [–0.20, –0.001], p_{LASSO_MCMC} = 0.047, Fig. 2H).

Clustering analysis of cognitive computation-related oscillatory activity

The preceding results show that the same IC (IC #6) correlated with oscillatory activity in opposite directions, depending on the underlying cognitive computation. This could suggest the existence of distinct deficit profiles among post-COVID-19 patients. To further investigate potential neurophysiological subtypes within this population, we conducted a model-based clustering analysis using a Gaussian Mixture Model (GMM). To determine the optimal number of clusters, we used the Bayesian Information Criterion (BIC). The clustering was based on standardized measures of three oscillatory features associated with cognitive computations, derived from the three cognitive tasks. We also included the diagnosis of COVID (yes/no) in the clustering procedure, not to differentiate post-COVID-19 patients from those with other respiratory diseases, but rather to identify subtypes within the COVID-19 population. In this context, including the COVID variable served as a control, allowing us to evaluate the clustering structure while considering the possibility that some non-COVID participants might share features with post-COVID-19 patients, and vice versa.

This analysis revealed three distinct clusters (Fig. 3A, B). Cluster 3 consisted almost exclusively of participants without a history of COVID-19, with only four exceptions. This suggests that individuals with no prior COVID-19 infection form a relatively homogeneous neurophysiological group. In contrast, the other two clusters (Clusters 1 and 2) included only participants recovered from COVID-19, and these were further differentiated based on their oscillatory pattern.

Fig. 3

Model-based cluster analysis revealed distinct patterns of oscillatory activity associated with cognitive computation in individuals who have recovered from COVID-19. (A) The clustering analysis identified three distinct patient subgroups. Cluster 3 included only individuals without a history of COVID-19, except for four cases. Clusters 1 and 2 comprised exclusively patients who had recovered from COVID-19. The x and y axes represent the first two principal components, which capture the majority of the variance in the four variables used in the clustering: COVID-19 diagnosis and oscillatory activity associated with three cognitive computational processes. (B) This panel depicts only Clusters 1 and 2, corresponding to post-COVID-19 patients. The x and y axes represent the first two principal components derived from the three oscillatory activity variables. The black line represents the decision boundary of a logistic regression model fitted to the two clusters, illustrating their separability in the reduced feature space. (C) Posterior distribution of the difference in mean theta-band oscillatory activity between patients in Cluster 1 and Cluster 2, related to conflict expectation. The p-value represents the posterior probability derived from the Markov Chain Monte Carlo (MCMC) sampling. (D) Posterior distribution of the difference in mean theta-band oscillatory activity between patients in Cluster 1 and Cluster 2, related to memory load. The p-value represents the posterior probability derived from the MCMC sampling. (E) Posterior distribution of the difference in mean theta-band oscillatory activity between patients in Cluster 1 and Cluster 2, related to uncertainty processing during decision making. The p-value represents the posterior probability derived from the MCMC sampling.

Within the post-COVID-19 group, Cluster 1 (compared to Cluster 2) showed preserved theta activity related to conflict expectation (Cluster 1 median = 0.29, p = 0.002, r = 0.59; Cluster 2 median = 0.07, r = 0.19, p = 0.2; between-cluster difference, r = 0.29, p = 0.019, Wilcoxon test, Bayesian posterior distribution of difference, median = 0.19, HDI = [ 0.05 0.34], p_MCMC = 0.005, Fig. 3C). In contrast, Cluster 1 exhibited altered oscillatory activity related to memory load (Cluster 1 median = 0.01, r = 0.02, p = 0.9; Cluster 2 median = 0.41, r = 0.68, p = 0.00002; between-cluster difference r = 0.48, p = 0.0001, Wilcoxon test, Bayesian posterior distribution of difference, median = -0.49, HDI = [−0.74 −0.24], p_MCMC = 0.0003, Fig. 3D), and uncertainty during decision-making (Cluster 1 median = 0.19, r = 0.53, p = 0.007; Cluster 2 median = −0.44, r = 0.77, p = 5 × 10⁻⁸; between-cluster difference r = 0.63, p = 5 × 10⁻⁶, Wilcoxon test, Bayesian posterior distribution of difference, median = 0.74, HDI = [ 0.51 0.99], p_MCMC < 0.0001, Fig. 3E). Notably, the increase in activity observed in Cluster 1 during uncertainty is paradoxical, as the overall population trend is characterized by a decrease in oscillatory power under uncertainty. Cluster 1 comprised 66% of patients with anosmia and 41% of those who had been hospitalized, whereas Cluster 2 included 50% of patients with anosmia and 45% with a history of hospitalization. However, these differences were not statistically significant (anosmia: χ² = 1.00, p = 0.29; hospitalization: χ² = 0.01, p = 0.90). The clusters show no difference in age (Wilcoxon test, p = 0.15), time between diagnosis and evaluation (p = 0.46), or sex (χ² = 1.9, p = 0.16).

We then evaluated whether the ICs identified from oscillatory activity could differentiate patient clusters. Interestingly, IC#4 distinguished between clusters (Wilcoxon test r = 0.28, p = 0.04; Bayesian logistic regression: median β = −0.57, 95% HDI = [−1.20, −0.01], p_MCMC = 0.04), whereas IC#6 did not show a significant difference (Wilcoxon test r = 0.12, p = 0.32; Bayesian logistic regression: median β = 0.40, 95% HDI = [−0.20, 0.60], p_MCMC = 0.19). At the behavioral level, clusters did not differ in the expectation of conflict stimuli (β = 0.01, SE = 0.015, t = −0.9, r = -0.09, p = 0.30). However, cluster 1 showed reduced performance under increasing memory load (β = -0.10, SE = 0.05, t = -2.18, r = −0.23, p = 0.029) and a lower learning rate in response to negative feedback (β = -0.41, SE = 0.17, t = -2.3, r = −0.24, p = 0.01).

These findings indicate that oscillatory brain activity patterns can differentiate not only between individuals with and without a history of COVID-19, but also among those who have had the disease, suggesting the existence of neurophysiological subtypes within the post-COVID-19 population. These subtypes may be crucial for guiding targeted, neurobiologically informed interventions.

Clustering robustness analysis

To assess the robustness of the clustering solution, we conducted several validation analyses following current best practices for clustering in neural data⁴⁸. First, we re-estimated the model on 1,000 bootstrap subsamples (each comprising 90% of the data) while adding Gaussian noise to all observations. From these iterations, we derived a consensus matrix and quantified its agreement with the original clustering. The resulting stability was high (median Adjusted Rand Index [ARI] = 0.80 for a 50% consensus threshold, and 0.74 for a 70% threshold), indicating that the two-cluster solution remained consistent across perturbations.

Additionally, we performed permutation-based control analyses to verify that the observed clusters reflected genuine data structure rather than random partitions. When diagnostic labels were permuted 1,000 times, the resulting partitions showed negligible similarity to the original one (median ARI = 0.06, interquartile range (IQR) [− 0.02, 0.10]). Permuting only the EEG-derived features yielded slightly higher, but still low, concordance (median ARI = 0.20, IQR [0.10, 0.30]), further supporting that the identified clusters are not driven by noise or single-modality artifacts.

Finally, to rule out the possibility that clustering reflected inherent noise in the EEG signals, we compared noise-related parameters between clusters. Specifically, we analyzed the percentage of rejected trials and the proportion of interpolated electrodes per trial. None of these measures differed significantly between clusters, confirming that the clustering results were not influenced by differences in EEG data quality (Rejected trials, KW test, statistic = 1.1, df = 2, p = 0.5; Interpolated channels, statistic = 0.06, p = 0.96).

Discussion

Our findings demonstrate that SARS‑CoV‑2 infection is associated with disruptions in neural oscillatory activity and white matter integrity that relate to specific cognitive profiles. While previous EEG studies have described electrophysiological alterations in recovered COVID‑19 patients at rest⁴⁹, our results extend this evidence by revealing changes in slow‑wave oscillatory patterns that reflect disruptions in specific neural circuits and cognitive computations. Specifically, the reduced alpha oscillations we observed in COVID‑19 patients diverge from the causal role of alpha activity in modulating spatial attention and visual processing documented in healthy individuals²⁴, suggesting impaired attentional suppression mechanisms. Similarly, the altered delta/theta oscillations in prefrontal cortex align with known default communication patterns between prefrontal and visual regions identified in healthy subjects²⁸ indicating a disruption of coordinated top-down control processes. These functional alterations are directly linked to changes in white matter microstructure, consistent with previous reports of reduced fractional anisotropy and increased radial diffusivity in individuals recovered from COVID‑19^8,22. Such white matter abnormalities have been associated with episodic memory deficits, attentional difficulties, and reduced verbal fluency, yet prior studies have not delineated how these structural changes map onto specific cognitive profiles. Although proposed mechanisms include direct viral invasion of neural tissue⁵⁰, immune‑mediated neuroinflammation⁵¹, and disruption of protective barriers^52,53, there has been limited evidence clarifying which neural circuits are most affected and how this heterogeneity manifests across patients. By linking oscillatory dynamics, white matter integrity, and computational cognitive profiles, our study provides critical evidence that these alterations reflect identifiable mechanisms that may explain the variability of cognitive outcomes among COVID‑19 survivors.

The cognitive profiles identified in our study align with previous reports of deficits in individuals with a history of SARS‑CoV‑2 infection, independent of the severity of acute respiratory symptoms and anosmia⁴. The alterations observed in low‑frequency oscillatory activity across multiple cognitive tasks indicate a disruption of neural synchronization mechanisms that depend on the integrity of long‑range white matter pathways⁵⁴. Importantly, these electrophysiological changes are directly associated with microstructural alterations in specific tracts, including the cingulum bundle and thalamocortical pathways—particularly the parietal and occipital fasciculi. This structure–function coupling provides a plausible neural substrate for the heterogeneous cognitive symptoms frequently reported in post‑COVID-19 populations and supports the idea that distinct network disruptions underlie specific patterns of cognitive impairment.

A key contribution of our study is the identification of two distinct neurocognitive profiles among individuals recovered from COVID‑19. One profile (Cluster 1) is characterized by working memory deficits, with oscillatory activity correlating with the integrity of the cingulum, consistent with disruption of the frontoparietal network functioning, which plays a central role in memory maintenance^41,55. Notably, these reductions in frontal theta oscillations in this profile represent a departure from the theta-mediated coordination between prefrontal and parietal regions that underlies normal working memory processes^28,56. This profile also exhibits marked impairments in decision‑making under uncertainty, with oscillatory activity correlating with the integrity of the thalamo‑occipital and uncinate fasciculi, suggesting dysfunction in decision‑making processes and specific forms of learning and memory⁵⁷. In contrast, the second profile (Cluster 2) showed preserved working memory and decision‑making performance but exhibited reduced computation of conflict expectation, a key mechanism of proactive cognitive control³⁵. Such cognitive computations are typically generated by theta oscillations in frontal and parietal regions, consistent with the role of theta synchronization in maintaining proactive control mechanisms²⁸; therefore, the observed reduction in theta-based conflict computation in this profile indicates a selective disruption of prefrontal–parietal theta coordination, a mechanism crucial for anticipating and preparing responses to cognitive demands in healthy subjects. Variation among patients in this theta-mediated process correlates with the integrity of the thalamus-parietal fasciculum. The identification of these distinct profiles, together with the observed variability in structure–function relationships, underscores the need for reliable biomarkers to capture this heterogeneity and guide the development of personalized assessment and intervention strategies³⁹.

Several limitations should be considered when interpreting our findings. The cross‑sectional design limits causal inference and does not allow assessment of the long‑term trajectory of the observed alterations. Although we controlled for multiple potential confounders, unmeasured factors—such as subclinical hypoxia or systemic inflammation—may also have contributed to the results. Furthermore, the moderate sample size may limit the generalizability of our findings, especially for subgroup analyses aimed at characterizing specific neurocognitive profiles.

Future research should adopt longitudinal designs to determine whether the neural alterations we identified reflect transient maladaptive responses or permanent damage. Targeted interventions, such as non-invasive brain stimulation (NIBS) tailored to specific oscillatory deficits, represent a promising therapeutic avenue. For instance, low-frequency repetitive transcranial magnetic stimulation (rTMS) applied to the dorsolateral prefrontal cortex (DLPFC) has been shown to improve cognitive function⁵⁸ and reduce depressive symptoms, fatigue, and apathy⁵⁹, although its efficacy depends on stimulation parameters⁶⁰. Moreover, more precise transcranial magnetic stimulation (TMS) aimed to enhance specific oscillatory activity can induce endogenous oscillations, offering the potential for more personalized and effective interventions^35,61. Combining cognitive training with NIBS—particularly in well-defined subgroups⁶²—has also shown promise in enhancing neural efficiency and cognitive performance⁶³. While preliminary results are encouraging, further large-scale controlled trials are essential to refine stimulation protocols and establish treatment efficacy. In this context, our findings provide critical insights into specific cognitive profiles, white matter tract integrity, and related functional parameters, which may guide intervention strategies. Furthermore, they underscore the potential of multimodal neuroimaging biomarkers for early identification of individuals at risk for persistent post-COVID-19 cognitive impairment.

In conclusion, this study provides evidence that SARS‑CoV‑2 infection can lead to disruptions in brain function and structure, even in individuals who experienced mild acute illness^64,65. By identifying distinct oscillatory patterns linked to specific neurocognitive profiles, our findings suggest that post‑COVID-19 cognitive changes are not a single syndrome but encompass multiple trajectories of neural dysfunction⁶⁶. These neurophysiological profiles may represent early evidence of distinct post-COVID-19 neurocognitive phenotypes, challenging the notion of a unitary “post-COVID-19” syndrome. Rather than relying on binary clinical classifications (e.g., post-COVID-19 vs. non-COVID-19), our approach highlights the value of mapping dimensional trajectories and intermediate neuro-cognitive mechanisms. This framework offers a scalable and mechanistically informed model that could be extended to other post-infectious or neurodegenerative conditions. These results also have important implications for the clinical management of post‑COVID-19 patients and contribute to a broader understanding of how viral infections can affect brain networks and cognitive processes³⁰. Importantly, these effects are not confined to traditionally vulnerable groups^66,67, as cognitive alterations were also observed in individuals without typical risk factors^68,69. The emerging evidence highlights the urgent need for ongoing research to track the long‑term neurological impacts of COVID‑19 and to develop early, targeted interventions aimed at preventing the progression of cognitive deficits, particularly as these effects may interact with age‑related neurocognitive decline.

Methods

Sample

We invited participants from a previously described cohort of 100 adult patients with respiratory infections, recruited between February 2020 and May 2023 from both public and private hospitals in Santiago, Chile, to a follow-up electroencephalography (EEG) session. Enrollment occurred 2–27 months post-diagnosis. From this cohort, 96 participants were evaluated and used throughout all analyses reported in this study. Participants ranged in age from 19 to 66 years (median = 39, mean = 40.1) and were evaluated after recovery from infection.

The cohort included individuals with mild to moderate respiratory symptoms, according to established severity classification criteria^70,71,72, who did not present respiratory failure nor require invasive mechanical ventilation. All participants were asymptomatic for respiratory symptoms for at least four weeks before evaluation. Of the total sample, 70 patients had COVID-19 confirmed by SARS-CoV-2 infection via PCR testing. The remaining 26 patients had respiratory infections attributed to other pathogens, confirmed by at least two negative SARS-CoV-2 PCR tests during the acute phase or follow-up, and no history of COVID-19-related symptoms or positive test results at any time.

There were no significant differences in age between the post-COVID-19 and non-COVID-19 groups (Wilcoxon test, p = 0.11, r = 0.16). Exclusion criteria included any history of intensive care unit (ICU) admission or invasive ventilation, pre-existing brain lesions, neuropsychiatric disorders, vascular events before or during the infection, or acute neurological symptoms beyond anosmia (e.g., seizures).

Cognitive, emotional, olfactory, and functional assessment

All participants from the previously described cohort underwent a comprehensive assessment battery to evaluate cognitive functioning, emotional symptoms, olfactory performance, and physical capacity. The following standardized instruments were applied: 1) Executive Functioning was assessed using the Chilean version of the INECO Frontal Screening (IFS-Ch)⁷³, which evaluates response inhibition, set-shifting, working memory, and abstraction capacity; 2) Global Cognitive Performance was measured using the Chilean version of the Addenbrooke’s Cognitive Examination III (ACE-III)⁷⁴, which includes five subscales covering orientation and attention, memory, verbal fluency, language, and visuospatial skills; 3) Anxiety Symptoms were assessed using the Spanish version of the Generalized Anxiety Disorder 7-item scale (GAD-7)⁷⁵, a brief self-report instrument composed of seven items aligned with DSM-IV criteria for Generalized Anxiety Disorder; 4) Depressive Symptoms were measured using the Spanish version of the Patient Health Questionnaire-9 (PHQ-9))⁷⁶, which includes nine items assessing the presence and severity of depressive symptoms over the previous two weeks, based on DSM-IV criteria; 5) Olfactory Function was evaluated with the KOR test⁷⁷, a validated screening tool for COVID-19–related olfactory deficits. Participants were presented with six easily distinguishable scents (e.g., mint, peach, onion), each of which had to be matched with images shown on an online platform. A score below 4 points indicated suspected olfactory dysfunction; finally, 6) Functional physical Capacity was assessed using the Six Minute Walk Test (6MWT)⁷⁸, a validated measure of submaximal aerobic capacity and endurance in clinical populations.

Experimental tasks

During the experimental session participants completed three behavioral tasks while their brain activity was recorded with electroencephalography. The Go-Nogo (GNG) task^35,42,79 was used to assess inhibitory control and conflict expectation. Visual stimuli consisting of the letters “X” and “O” were presented centrally on a screen for 300 ms each, with an interstimulus interval randomly varying between 900 and 1400 ms. The task was divided into two blocks of 150 trials. For each participant, one letter was assigned as the Go stimulus and the other as the Nogo stimulus in the first block, with the assignment reversed in the second block. Stimulus assignment was randomized and counterbalanced across participants. The occurrence of Nogo stimuli was fixed at 25% (i.e., a Nogo-to-Go ratio of 1:3), and the number of consecutive Go trials preceding a Nogo trial was not constrained. This design resulted in a variable distribution of Go-sequence lengths, with the frequency of sequences ranging from approximately p = 0.18 for a single Go trial before a Nogo to p < 0.10 for sequences longer than seven Go trials. This structure allowed for a broad sampling of expectancy and conflict levels across trials and participants.

A modified version of Sternberg’s Memory Scanning Task was used to assess working memory, in particular aiming to manipulate working memory (WM) load, following the procedures described in detail by Figueroa-Vargas et al⁴¹. In each trial, participants encoded a memory set of 2, 4, or 6 consonants arranged circularly around a fixation cross that were presented for 1.8 s, followed by a maintenance period of 2 s, and a probe stimulus requiring a yes/no response. Each participant completed 270 trials across randomized blocks.

The Reversal Learning Task was used to assess decision-making under uncertainty. Participants repeatedly chose between two decks of cards, each associated with a potential score. After each selection, they received feedback indicating whether the chosen deck paid the indicated reward. The reward probabilities for each deck were either 0.85/0.15 or 0.75/0.25 and reversed unpredictably every 10–15 trials. Each participant completed four games per session. This task required participants to learn and update reward contingencies through trial and error in a volatile environment, fostering adaptive decision-making under uncertainty. For further task details, see our previous work⁸.

For all tasks, stimulus presentation and response collection were controlled using Presentation® software (version 21.1, Neurobehavioral Systems, https://neurobs.com/).

Behavioral analysis

For the GNG task, we first estimated participants’ sensitivity to No-Go stimuli using the d′ (d-prime) index from signal detection theory. For each participant, we computed the hit rate (HR = hits / [hits + misses]) and false alarm rate (FAR = false alarms / [false alarms + correct rejections]). To avoid infinite values, extreme rates of 0 or 1 were adjusted using a standard correction (HR = 1 − 1/(2N), where N is the number of Go or No-Go trials). d′ was then calculated as the difference between the z-scores of the hit and false alarm rates: d′ = z(HR) − z(FAR). To examine how Go sequence history affected performance, trials were classified as ‘short’ (< 3 consecutive Go trials) or ‘long’ (3–7 consecutive Go trials), including only those with reaction times between 40 and 1000 ms. For each participant, mean error rates and reaction times were computed for both conditions. Paired Wilcoxon and t-tests were used to assess the effect of sequence length, and a linear mixed-effects model tested interactions with experimental factors. Then, expectative of conflict was calculated using a Drift Diffusion Model (DDM) in which the decision boundary parameter was modulated by trial-wise expectation of conflict. Conflict expectation (Ex) was derived from both the sequential structure of stimuli and participants’ ongoing learning about the probability of encountering Nogo trials. Specifically, the model assumes that participants estimate the probability of conflict (Q) using a Rescorla-Wagner learning rule: Q_t = Q_t−1 + α(C_t−1 − Q_t−1), where α is the learning rate and C_t−1 indicates whether the previous trial was a Nogo (C_t−1 = 1) or Go (C_t−1 = 0) trial. This estimated Q was then combined with a sequence-based transformation reflecting the impossibility of two consecutive No-Go trials, generating a dynamic expectation value that increased early in Go sequences and decreased in longer ones: Ex = 1-(1-Q)^(Seq-1). See more details and model validation in Martínez-Molina et al³⁵.

For the WM task, we analyzed response accuracy using hierarchical logistic models. Predictors included memory load (i.e., the number of letters to be memorized on each trial) and whether the probe item was present in the original memory array, as reported in prior work^41,45,46,80.

For the RL task, we analyzed participants’ choices using the same computational cognitive modeling approach described in detail in our previous work⁸. In brief, models were based on Prospect Theory, in which the subjective value of each option was computed by combining a value function with subjective probability weighting. To estimate the unknown probabilities of receiving a reward, we implemented a Rescorla-Wagner learning algorithm, where probabilities were updated trial by trial based on the discrepancy between expected and received outcomes, as follows: p_t = p_t-1 + α(Feedback_t-1 − p_t-1). Here, p_t represents the estimated probability at trial t, α is the learning rate, and Feedback_t-1 is a dummy-coded variable indicating whether the previous choice was rewarded (1) or not (0). We estimated separate learning rates for positive and negative feedback, and included counterfactual updating—i.e., updating both the chosen and unchosen options, assuming the unchosen option would have yielded the opposite outcome. Choice behavior was modeled using a logistic function based on the subjective value difference between options, with a temperature parameter controlling decision stochasticity and a bias term capturing side preference.

For all computational cognitive models, learning parameters were estimated separately for each participant using Bayesian inference implemented in JAGS (version 4.2.0, https://mcmc-jags.sourceforge.io/) and R (version 4.2.1, https://www.r-project.org/) via MCMC sampling. Subsequently, the remaining model parameters were estimated using the full dataset within a hierarchical modeling framework. This approach allowed us to obtain population-level estimates for key model parameters while also enabling more computationally costly estimation procedures at the individual level. A similar approach has been used in our previous work, showing no bias in the resulting estimates^8,35,37.

EEG analysis

Brain activity was recorded using a 64-channel EEG system (BrainVision amplifier system, BrainProducts, Germany) with scalp electrodes arranged according to the international 10–10 system, using EGI Wet-Net caps. Signals were sampled at 1,000 Hz, bandpass filtered between 0.15–500 Hz during acquisition, and electrode impedance was kept below 5 kΩ. EEG data were analyzed with the LAN toolbox (version 1.10, https://github.com/neurocics/LAN_current) for MATLAB (version R2022a, https://www.mathworks.com/ ). Data were segmented into epochs centered around the stimulus of interest for each task: from − 1.5 to 1.5 s relative to Go stimulus onset in the Go/Nogo task, from − 1.5 to 4.5 s relative to the onset of the letter array in the Working Memory task, and from − 1.5 to 2.5 s relative to the presentation of options in the Reversal Learning task. All epochs were inspected for artifacts using a semi-automated pipeline^43,44,47,81. Artifacts were automatically detected using the following criteria: voltage amplitudes exceeding ± 100 μV, amplitude deviations greater than 3 standard deviations from the mean, spectral power exceeding 2 standard deviations across more than 10% of the 1–30 Hz frequency range, or abnormally low correlations between neighboring electrodes (i.e., below − 3 standard deviations). Ocular artifacts such as blinks were identified and removed using Independent Component Analysis (ICA). Trials with automatically detected artifacts that were confirmed by visual inspection were excluded. The remaining artifact-free data were re-referenced to the average reference and transformed using a surface Laplacian to estimate current source density. Data were then digitally bandpass filtered between 0.1 and 45 Hz. Time–frequency decomposition was performed using a 5-cycle Morlet wavelet transform to compute the spectral power distribution. The time windows for this analysis ranged from − 0.5 to 1.0 s for the Go/Nogo task, − 0.5 to 3.8 s for the Working Memory task, and − 0.5 to 1.5 s for the Reversal Learning task, all relative to the onset of the task-relevant stimulus.

EEG statistical analysis

For the EEG statistical analysis, we first fitted a general linear model (GLM) of the power of the oscillatory activity per trial in each participant [first-level analysis, (see 51–55)] using the following equation:

$${\text{GoNogo task}}:{\text{ Power}}\left( {{\text{f}},{\text{t}}} \right) \, \sim \beta_{{1}} + \beta_{{2}} {\text{Nogo }} + \beta_{{3}} {\text{Ex}}*{\text{Go }} + \beta_{{4}} {\text{Ex}}*{\text{Nogo}}$$

(1)

$${\text{WM task}}:{\text{ Power}}\left( {{\text{f}},{\text{t}}} \right) \, \sim \beta_{{1}} + \beta_{{2}} {\text{ML }} + \beta_{{3}} {\text{SMP }} + \beta_{{4}} {\text{ML}}*{\text{SMP}}$$

(2)

$${\text{RL task}}:{\text{ Power}}\left( {{\text{f}},{\text{t}}} \right) \, \sim \beta_{{1}} + \beta_{{2}} {\text{V }} + \beta_{{3}} {\text{U}}$$

(3)

In the preceding equations, the variation in the power of the oscillatory activity as a function of frequency (f) and trial time (t) depends on a series of variables specific to each task, where β₁ represents the intercept and β₂ to β₄ are the slopes (coefficients) of the corresponding variables. For the GoNogo task, Nogo represents the Nogo stimuli, Go the Go stimuli, and Ex the theta latent variable derived from computational cognitive modeling, reflecting the expectation of conflict. For the WM task, ML refers to memory load, taking the values 2, 4, or 6; while SMP denotes successful memory performance on the current trial. Finally, for the RL task, V represents the value of the chosen option, and U the uncertainty of the trial, derived from the previous trial—both extracted from computational cognitive modeling.

We then obtained a 2D matrix of normalized β-values (time × frequency), computed as β-value divided by its standard error, for each regressor and participant within the selected ROI of electrodes. To assess group differences under selected conditions, we first conducted a Kruskal–Wallis test (second-level analysis). We then applied a cluster-based permutation (CBP) test to correct for multiple comparisons. Briefly, in this method, clusters of significant effects were defined by grouping neighboring time–frequency points and adjacent electrodes that exhibited the same statistical effect (cluster-threshold detection, CTD; uncorrected p < 0.05). For each cluster, a cluster-level statistic was calculated as the sum of the test statistics across all sites within that cluster (e.g., Z-values for the Wilcoxon test). The significance of each observed cluster was then evaluated using a permutation distribution built from the largest cluster-level statistic obtained in each iteration. This distribution was generated by randomly permuting the original data—either by shuffling regressor labels across trials (within-subject analysis) or group labels (between-subject analysis). For each permutation, the original statistical test (i.e., Wilcoxon or Kruskal–Wallis) was recomputed, and the largest cluster-level statistic was retained. After 5,000 permutations, the p-value of each observed cluster was calculated as the proportion of permutations in which the cluster-level statistic exceeded that of the observed cluster.

Approximation for Cluster-level effect size. Following the cluster-based permutation framework, no canonical cluster-level effect size is defined. Therefore, we report an approximate magnitude derived from the underlying mass-univariate statistic. Specifically, for each significant cluster, we (i) converted the per-pixel test statistic to a standardized effect metric (e.g., Cohen’s d / Hedges’ g or standardized β), and (ii) summarized the cluster magnitude as the average across the top-K suprathreshold pixels, with K set to min(200, 0.2·cluster size). This summary provides an interpretable, standardized approximation of cluster magnitude while the cluster-level inference (significance) remains determined by the permutation test.

Source analysis

We used Brainstorm (version 23–02-2023, https://neuroimage.usc.edu/brainstorm) software to estimate time-resolved neural current density at the cortical surface by applying a minimum norm estimate (MNE) inverse solution based on the sLORETA algorithm, with unconstrained dipole orientations. We used a tessellated cortical mesh to model the brain based on each individual’s anatomy. This mesh was employed to estimate the distribution of current sources, with ~ 3 × 10,000 sources positioned on the segmented cortical surface (three orthogonal sources at each spatial location). Source reconstruction was performed on single-trial EEG data for each condition and subject. To compute the forward model, we employed a multi-layer continuous Galerkin finite element conductivity model using SIMNIBS (version 4.1.0, https://simnibs.github.io/simnibs) and DuNEURO (version 240,107, https://www.medizin.uni-muenster.de/duneuro) software. The source current was estimated by multiplying the recorded EEG time series by the inverse operator. For time–frequency analysis, the three dipole components at each source location were reduced to a single value by projecting them along the direction of maximal variance for each trial. Time–frequency decomposition was then performed directly in source space using Morlet wavelet transforms.

DTI analysis

Diffusion MRI data were acquired during a previous MRI session of the participants (REF). Data were processed using DSI Studio (version 2023.10, http://dsi-studio.labsolver.org/), which includes eddy current correction, motion correction, and diffusion tensor model estimation. From this model, the fractional anisotropy (FA) diffusion metrics were extracted. Whole-brain fiber tractography was performed using a deterministic tracking algorithm⁸², with the following parameters: angular threshold = 60°, step size = 1 mm, minimum tract length = 30 mm, maximum tract length = 250 mm, smoothing = 0.5, and QA threshold = 0. Fiber bundle segmentation was carried out using the Deep White Matter (DWM) bundle atlas⁸³, which includes 36 well-characterized major tracts. Automatic segmentation was performed using an algorithm based on the maximum Euclidean distance between corresponding points of fiber trajectories⁸⁴. See more details in^8,38.

EEG-DTI analysis

In the first step, we conducted an independent component analysis (ICA) to study the relationship between FA and oscillatory activity. ICA allowed us to reduce the complexity of DTI data while retaining the most variance, facilitating an exploration of the associations between brain structures and fatigue scores across the study groups. This approach has been used for complex data from MRI and integrated modalities by other groups⁸⁵. Additionally, as ICA generates independent components, the resulting components can be effectively utilized in linear regression analyses such as Bayesian Least Absolute Shrinkage and Selection Operator (LASSO) analysis. This approach is particularly advantageous because LASSO regression helps control for multicollinearity among predictors, ensuring the model remains robust and interpretable (see above). Importantly, this two-step strategy—dimensionality reduction through ICA followed by Bayesian LASSO regression—acts as a model-based correction for multiple comparisons. By representing the DTI data with a limited set of orthogonal components, the number of effective statistical tests is greatly reduced, and by applying an L₁-penalized regression, the model automatically shrinks non-informative predictors toward zero, minimizing false-positive associations. Together, the ICA–Bayesian LASSO pipeline provides an integrated and data-driven approach that effectively controls for both multicollinearity and Type I error inflation across multimodal EEG–DTI analyses. Moreover, using ICA-derived components allows for multiple comparison corrections, reducing the risk of Type I errors⁸⁶.

In the second step, we applied Bayesian Least Absolute Shrinkage and Selection Operator (LASSO) linear regression models using R in conjunction with the Just Another Gibbs Sampler (JAGS: A program for analysis of Bayesian graphical models using Gibbs sampling). LASSO offers robust frameworks for addressing the challenges associated with multiple comparisons in regression analysis. Their ability to perform automatic feature selection while controlling for overfitting makes them ideal tools for high-dimensional data analysis. This method enabled the analysis of compositional data, improving the accuracy of parameter estimation and offering more profound insights into the relationships between the compositional predictors—namely, the ICA of FA per track—and the outcome variable of oscillatory activity. Since compositional data inherently reflect the relative nature of its components, analyzing them allows for better differentiation of ratings and reduces response biases. This approach is effective whether all components contributing to the total are fully quantified or only a subset is included⁸⁷. Thus, these methods mitigate the type I error rate through its inherent feature selection process, as it tends to include only a subset of statistically significant predictors, thus reducing the number of tests conducted on non-informative variables⁸⁸. For these analyses, we used the first 20 components that represent the 90% of variance of the data.

For LASSO, we used the following statistical model:

$${\text{Power }}\sim \beta_{0} + \beta_{{{1},{\text{i}}}} {\text{IC}}_{{\text{i}}} + \beta_{{2}} {\text{Co}} + \beta_{{3}} {\text{An}} + \beta_{{4}} {\text{H}} + \beta_{{5}} {\text{Age}} + \beta_{{6}} {\text{Time}}$$

(4)

In this equation, β₀ represents the intercept, while β_1,i represents the slope corresponding to the independent component i (i = 1 to 20). The term β₃ to β₆ represents the slope corresponding to clinical variables, Co: diagnosis of COVID-19, An: Anosmia during acute episode, H: hospitalization during acute episode, Age: age of the patient, Time: time between diagnosis and experimental session. Inferences for significant regressors were assessed using the 95% highest density interval (HDI) of the posterior distribution. A p-value equivalent (p_{LASSO_MCMC}) was calculated by comparing Markov Chain Monte Carlo (MCMC) samples against a reference value of zero. To identify the fascicles that most strongly influenced each model, we first selected components whose 90% HDI did not include zero (equivalent to p_MCMC < 0.1). For each of the selected components, we multiplied the posterior distribution of the corresponding beta coefficient by the tract-wise weights associated with that component. These weighted distributions were then summed across components to obtain a posterior weight distribution for each fascicle. To identify relevant fascicles, we selected those whose posterior weight distribution had a 95% HDI entirely greater than zero. An exception was made for the RL model, where negative effects were expected; in those cases, fascicles with a 95% HDI entirely below zero were selected. This approach accounts for potential trade-offs between components that may exert opposing effects on the same fascicle. It allowed us to highlight consistent effects across the most relevant components identified by the Bayesian LASSO regression.

P-values below 0.05 were considered statistically significant, and all comparisons were two-tailed.

Gaussian mixture model-based clustering analysis

To identify potential neurophysiological subtypes among participants, we applied a model-based clustering approach using Gaussian Mixture Models (GMMs) as implemented in the mclust R package⁸⁹. This method probabilistically assigns individuals to latent clusters based on their multivariate feature profiles, assuming that the data arise from a mixture of Gaussian distributions. Parameters were estimated via the Expectation–Maximization (EM) algorithm. The feature set used for clustering consisted of standardized oscillatory brain activity measures obtained from three cognitive computations: expectation of conflict, memory load, and uncertainty during decision making, as well as a binary clinical variable indicating a history of COVID-19 infection (0 = no, 1 = yes). All continuous variables were z-scored prior to analysis. Importantly, the inclusion of the COVID-19 diagnostic variable in the clustering model was not intended to discriminate COVID vs. non-COVID participants. Instead, it was introduced to ensure that the resulting clusters would account for the effect of diagnosis and allow the identification of meaningful neurophysiological subtypes within the spectrum of post-COVID-19 patients. In other words, the clustering aimed not to establish a diagnostic classifier, but to uncover heterogeneity among individuals who shared a COVID-19 diagnosis. We fitted GMMs with varying numbers of clusters and covariance structures, and selected the optimal model using the BIC. The model that maximized the BIC was selected, and each participant was assigned to the cluster with the highest posterior probability.

Cluster robustness analysis

To evaluate the stability and reliability of the identified clusters, we conducted several control analyses in accordance with current recommendations for clustering in neural data (Nakuci et al.⁴⁸). Cluster robustness was first assessed using a consensus-based approach: the model was re-estimated on 1,000 bootstrap subsamples, each containing 90% of the original data, with Gaussian noise added to all observations (mean = 0, SD = 0.05 on z-normalized values).

For each iteration, the cluster assignments were stored to build a consensus matrix, where each element represented the proportion of times that two participants were assigned to the same cluster across all iterations. This matrix captures the co-assignment stability between all pairs of participants. To identify stable relationships, we applied consensus thresholds ranging from 0.5 to 0.7, such that values above a given threshold indicated pairs that were co-assigned in at least 50–70% of the iterations. From these matrices, we derived consensus partitions and quantified their agreement with the original clustering using the Adjusted Rand Index (ARI). High ARI values across thresholds indicate that the observed clusters remain consistent under data resampling and perturbation. We further performed permutation-based validation analyses to determine whether the clustering reflected genuine data structure rather than random partitions. In these analyses, the diagnostic labels or specific feature sets (e.g., EEG-derived measures) were permuted 1,000 times, and the resulting partitions were compared to the original clustering using the ARI and the IQR of the obtained distributions. Finally, to control for potential artifacts related to data quality, we examined EEG noise parameters—including the percentage of rejected trials and the proportion of interpolated trials or electrodes per participant—across clusters. Differences were tested using the Kruskal–Wallis test. These procedures provide a model-free estimate of cluster reproducibility, guarding against overfitting or spurious partitioning and ensuring that the identified clusters reflect genuine structure in the data rather than random variation.