Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), now endemic, causes long coronavirus disease (COVID) or post-acute sequelae of SARS-CoV-2 infection (PASC), which has become a significant healthcare burden¹. Long COVID presents with various symptoms, including respiratory, neurological, cardiovascular, gastrointestinal, musculoskeletal, and psychological issues that can persist for >1 year in both hospitalized and non-hospitalized patients¹. Its pathogenesis likely involves multiple mechanisms¹, including tissue damage, viral reservoirs, autoimmunity, and persistent complement-mediated immunopathology associated with thromboinflammation^2,3,4.

Brain alterations may occur after COVID-19 infection^5,6,7,8,9,10. A UK Biobank magnetic resonance imaging (MRI) study revealed increased diffusion indices in the limbic regions and striatum and reduced cortical thickness in the orbitofrontal and parahippocampal areas, regardless of cognitive dysfunction⁵. Recent imaging studies have indicated distinct microstructural changes in long COVID^6,7, with network disruptions varying by symptom type, such as cognitive impairment, olfactory dysfunction, and fatigue^7,9. Brain changes may be more pronounced in individuals with cognitive impairment, marked by blood-brain barrier (BBB) disruption⁸, elevated brain injury markers⁹, and reduced gray matter volume¹⁰, even 1 year after the infection. However, whether these brain changes represent previous injury or are accompanied by ongoing neurodegenerative processes, as well as whether such processes are universal across PASC patients, particularly in individuals who experienced mild COVID-19, is unclear.

Advanced magnetic resonance imaging techniques may offer insights into neurodegeneration following SARS-CoV-2 infection beyond gray matter measurements. Iron, which is crucial for metabolism, myelin synthesis, and neurotransmitter production, can contribute to neurodegeneration when accumulated excessively¹¹. Paramagnetic susceptibility MRI assesses brain iron levels by detecting faster T2*-weighted signal decay¹², with increased iron deposition observed in conditions such as Alzheimer’s disease (AD)¹³. The choroid plexus (CP), forming the blood–CSF barrier and producing CSF, plays a role in waste clearance and regulates immune cell entry¹⁴. Blood–CSF barrier disruption may impair clearance, contributing to neuroinflammation and neurodegeneration⁸. Structural changes in the CP, such as enlargement, were observed in neurodegenerative diseases^15,16. The glymphatic system, responsible for clearing brain waste via cerebrospinal fluid (CSF), can be disrupted by neuroinflammation, leading to abnormal protein aggregation—a hallmark of neurodegeneration¹⁷. Diffusion tensor imaging analysis along the perivascular space (DTI-ALPS), which measures radial water diffusion at the lateral ventricles, shows reduced indices when glymphatic function is impaired, as observed in various neurodegenerative diseases^18,19. In addition to ALPS indices, other DTI-derived metrics such as mean diffusivity (MD), radial diffusivity (RD), and fractional anisotropy (FA) can provide complementary insights: increased MD reflects overall microstructural damage and increased water diffusivity, elevated RD is often associated with demyelination, and reduced FA indicates loss of axonal integrity and fiber organization²⁰. Brain imaging studies assessing iron deposition, CP alterations, and these DTI parameters may help delineate neurodegenerative processes in the brain after SARS-CoV-2 infection.

Ultrasensitive or high-throughput protein analysis can offer a crucial understanding of the pathogenesis of long-COVID cognitive impairment. Neurons and glial cells release essential proteins, which can be drained into the CSF via the glymphatic system and subsequently into systemic circulation, even without blood-brain barrier disruption²¹. This release increases with brain damage, especially with BBB damage, offering clues about the extent of damage and ongoing pathology. For instance, elevated neurofilament light (NfL) and glial fibrillary acidic protein (GFAP) levels reflect neuronal and astroglial damage^22,23,24, while the altered Aβ42/Aβ40 ratio and increased phosphorylated tau levels suggest AD pathology^25,26. In addition to brain-specific proteins, blood contains markers of systemic inflammation and metabolism. While previous studies on long-COVID cognitive impairment identified unique proteomic signatures, they mainly focused on inflammatory and vascular proteins^27,28,29,30. A more comprehensive proteomic analysis, incorporating diverse biological pathways, could offer deeper insights into the mechanisms underlying the changes in long-COVID.

Therefore, we employed advanced brain MRI and high-throughput proteomics in individuals at least 3 months after COVID-19 (mainly 1 year after COVID-19), with mostly non-hospitalized mild COVID-19. Participants were grouped into three categories: PASC with cognitive impairment (Cog-PASC), PASC without cognitive impairment (Other-PASC), and no significant PASC (NS-PASC). We hypothesized that cognitive dysfunction in long COVID would show distinct neurodegenerative changes. To investigate this, we analyzed cross-sectional data from the LOCOMOTIVE study (LOng Covid Outcome MOnitoring with Translational InVEstigations of blood and neuroimaging; cris.nih.go.kr, KCT0007848), an ongoing prospective, longitudinal study of long COVID. In the exploration cohort, the Cog-PASC group showed elevated astroglial damage-associated proteins, cortical thinning in the cingulate and insular cortices, increased paramagnetic susceptibility in the hippocampus, and enlarged choroid plexus volume. In a covariate-matched cohort based on age, sex, and education, cortical thinning and susceptibility in the cingulate persisted. Proteomic profiling revealed additional alterations in oxidative stress responses and synaptic function, consistent with neurodegenerative pathways. In the independent validation cohort, elevations in neuronal and astroglial damage-associated proteins, cortical thinning in the cingulate and insular cortices, and increased hippocampal susceptibility were reproduced, confirming the robustness of these alterations. Together, these findings suggest unique neurodegenerative processes in Cog-PASC not prominent in other-PASC patients even after mild COVID-19.

The LOCOMOTIVE cohort consists of two independent groups, the exploration and validation cohorts, divided according to enrollment dates (Fig. 1a). In this preliminary cross-sectional study, participants with confirmed COVID-19 and positive SARS-CoV-2 antibody tests were classified into three groups: Cog-PASC (with cognitive impairment), Other-PASC (without cognitive impairment but with other-PASC symptoms), and no significant PASC (NS-PASC). Brain MRI and blood biomarker assessments were conducted in the exploration cohort, with confirmatory analyses performed in the validation cohort.

In the exploration cohort, 169 participants were prospectively recruited, of whom 160 were included in the analysis. Among the included patients, 28 individuals were classified as Cog-PASC, 55 as Other-PASC, and 77 as NS-PASC (Table 1). Of these, 107/160 (66.9%) were female. Participants were enrolled a median of 12.0 months after infection [IQR 7.0–14.0], and MRI was performed at a median of 13.0 months [IQR 8.0–15.0] post-infection, with no significant differences among groups. Most participants had been infected during the Omicron-dominant period.

The Cog-PASC group was older and had a higher proportion of females than the NS-PASC group but fewer females than the Other-PASC group (Table 1). Participants in the Cog-PASC group also had the fewest years of education. Given the unequal distribution of age, sex, and education across groups, these variables were adjusted for in all subsequent comparative analyses. Most participants experienced mild COVID-19 infections, not requiring hospitalization. The Cog-PASC group had significantly lower Montreal Cognitive Assessment (MoCA) scores than the Other-PASC and NS-PASC groups, with a median MoCA score of 24. Meanwhile, Fatigue Assessment Scale (FAS) scores, indicating more severe fatigue with higher values, were higher in the Other-PASC group compared to the Cog-PASC group. However, PASC scores were similar between the two groups, both showing mild symptoms (median score: 11, with a significant symptom threshold of 12). Symptom prevalence was higher in both Cog-PASC and Other-PASC groups compared to NS-PASC, with similar symptom patterns between the two PASC groups except for more frequent post-exertional malaise in Other-PASC (Fig. 1b, c). Notably, brain fog was reported at comparable rates between the two groups, aligning with previous findings that self-reported brain fog does not reliably reflect objective cognitive impairment³¹. These findings suggest that individuals with Cog-PASC also experienced other-PASC symptoms at levels generally similar to those in the Other-PASC group.

The Cog-PASC group shows brain damage and neurodegenerative changes

We analyzed blood samples to investigate potential brain damage in the Cog-PASC group using ultrasensitive single-molecule array (Simoa) technology to measure biomarkers of astroglial (GFAP) and neuronal (NfL) damage^23,24, as well as AD-related markers such as β-amyloid (Aβ42/Aβ40 ratios) and tau (p-Tau181)^25,26. The Cog-PASC group exhibited significantly higher GFAP levels than the other groups but comparable levels of the NfL, Aβ42/Aβ40 ratio, and p-Tau181 (Fig. 2a–d). GFAP and NfL were significantly correlated with cognitive dysfunction, as measured by MoCA scores, while AD-related biomarkers were not; however, these correlations became non-significant after adjusting for age, sex, and education (Supplementary Fig. 1a–d). Notably, neither GFAP nor NfL correlated with fatigue (FAS scores) or other long-COVID symptoms (PASC scores), even in unadjusted analyses, suggesting a more specific association with cognitive function than with general PASC symptoms (Supplementary Fig. 1e–h).

We assessed cortical gray matter thickness and volume using automated MRI processing, including segmentation and surface-based morphometry (Fig. 2e, f). Cortical thickness was generally comparable across most regions, except for the cingulate, insular cortex, and parahippocampal gyrus, which were thinner in the Cog-PASC group compared to the Other-PASC group (Table 2 and Fig. 2g). Insular cortex thickness showed a significant association with total MoCA scores in regression analysis adjusted for age, sex, and education (Fig. 2h). Meanwhile, volumetric comparisons of cortical and subcortical regions showed no group differences (Table 2).

We evaluated iron deposition in cortical and deep gray matter using the χ-separation method³², which maps paramagnetic susceptibility (Fig. 2e, i). Susceptibility was generally negatively correlated with cortical thickness, with a particularly significant association in the temporal cortex (Fig. 2j). The Cog-PASC group exhibited higher paramagnetic susceptibility, indicating increased iron deposition, in the cingulate and occipital cortices as well as the hippocampus (Table 2 and Fig. 2k). Among these regions, susceptibility in the occipital cortex and hippocampus was significantly associated with total MoCA scores (Fig. 2l). When we evaluated associations between these brain changes (whole cortical thickness or paramagnetic susceptibility) and fatigue or other long-COVID symptoms, no significant correlations were identified (Supplementary Fig. 2a–d).

Next, we assessed the CP volume fraction (CPV fraction) using segmentation of 3D T1-weighted images and evaluated various MRI parameters from diffusion tensor imaging (DTI), including the DTI-ALPS index (Fig. 2m). The CPV fraction was significantly higher in the Cog-PASC group than in the Other-PASC group. The DTI-ALPS index was lower in the Cog-PASC group, and Tract-Based Spatial Statistics (TBSS) demonstrated reduced FA with higher MD and RD in major white matter tracts. However, none of these differences reached statistical significance (Table 2, Fig. 2n, Supplementary Fig. 3a, b). Regression analyses revealed no significant associations between CPV fraction or DTI-ALPS index and MoCA scores (Fig. 2o), suggesting these brain measures may not directly reflect cognitive function. Similarly, neither metric revealed significant correlations with other symptom scores, including FAS and PASC scores (Supplementary Fig. 2e–h).

Collectively, these findings suggest that the Cog-PASC group exhibits evidence of brain injury and neurodegenerative changes, characterized by elevated astroglial damage markers, reduced cortical thickness, iron accumulation, and increased choroid plexus volume, without notable alterations in AD-related biomarkers.

Covariate matching confirms brain alterations in Cog-PASC

Due to significant differences in age, sex, and education among the PASC groups, we further conducted a covariate matching analysis to adjust for these variables. Participants were matched 1:1 using nearest-neighbor matching without replacement, based on logistic regression models that included age, sex, and education as covariates. Following matching, baseline characteristics were balanced across the PASC groups, with the exception of MoCA, FAS, and PASC scores (Table 3). The Cog-PASC group exhibited greater cognitive impairment, the Other-PASC group showed higher fatigue scores, and PASC scores remained comparable between groups, consistent with the original cohort findings (Table 1).

In analyses of astroglial/neuronal injury and AD-related blood biomarkers, the Cog-PASC group showed a trend toward higher GFAP levels, although this did not reach statistical significance. No significant differences were observed in other blood biomarkers across groups (Supplementary Fig. 4a–d). Regarding brain MRI biomarkers, the Cog-PASC group demonstrated significantly thinner cortex and increased paramagnetic susceptibility in the cingulate cortex (Table 4). Similar trends toward increased paramagnetic susceptibility were observed in the occipital and insular cortices and hippocampus, though they did not achieve statistical significance. The CPV fraction and DTI-ALPS index did not significantly differ between groups.

Taken together, these findings suggest that evidence of brain damage and neurodegenerative changes on MRI persists even after covariate matching, replicating the results from the original group comparisons.

The Cog-PASC group has distinct blood proteomic signatures linked to neurodegeneration

To explore the mechanisms underlying Cog-PASC, we applied high-throughput Olink proteomics²⁹ to measure 3072 serum proteins. Although a principal component analysis was performed to differentiate the three groups, the separation was not distinct, except for the Cog-PASC group, which showed a non-overlapping region (Fig. 3a). No significant differences were observed between PASC (Cog-PASC and Other-PASC) and NS-PASC groups, reflecting the chronic and mild nature of PASC symptoms in this cohort (Supplementary Fig. 5a).

Comparison between Cog-PASC and NS-PASC groups revealed 44 differentially expressed proteins (DEPs), including 43 downregulated DEPs and one upregulated DEP (Fig. 3b and Supplementary Table 1)^{33,34,35,36,37}. The downregulated DEPs were functionally linked to synaptic regulation and neurotransmission (e.g., WASF3, ARHGAP1, ADD1, EIF4B, SH3GLB2), oxidative stress response (SOD1, CCS), de-ubiquitination (YOD1, STAMBP), and apoptotic signaling (CRADD). The five most prominently downregulated proteins were WASF3, YOD1, ARHGAP1, CA2, and CRADD. CASQ2 was the only upregulated protein associated with calcium ion regulation. No significant protein expression differences were observed between the Cog-PASC and Other-PASC groups or between Other-PASC and NS-PASC groups (Fig. 3c and Supplementary Fig. 5b).

Gene Ontology (GO) analysis of the DEPs highlighted pathways related to oxidative stress and synaptic functions, while Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis identified neurodegeneration as the most significant pathway (Fig. 3d, e). Protein–protein interaction analysis revealed four main clusters centered on SOD1, STAMBP, MRI1, and EIF4B (Fig. 3f). MoCA scores were significantly associated with 24 of the 44 DEPs (54.5%) in regression analysis, whereas no associations were found with FAS or PASC scores (Fig. 3g), suggesting a specific relationship between these proteins and cognitive function.

Then, we examined various proteins, including immune and vascular regulatory proteins, particularly those recently reported as key contributors to long COVID symptoms and neurodegeneration^3,8,38. Using the NS-PASC group as a reference, six DEPs (TREM2, C1QA, SPON-1, CXCL10, IL-4, and VCAM-1) showed significant differences in the Cog-PASC group, while four DEPs (C2, CD33, PDGF-β, and TNF) differed significantly in the Other-PASC group (Fig. 3h). Among these, TREM2, CXCL10, IL-4, VCAM-1, CD33, and TNF are inflammatory proteins^4,8,38. C2 and C1QA are part of the complement system³, and PDGF-β regulates pericytes⁸, maintaining vascular function⁴, including the BBB³⁹. SPON-1 is involved in neuronal regulation and cell adhesion⁴. All these proteins, except IL-4 and PDGF-β, were upregulated in the Cog-PASC group. The downregulation of IL-4 and PDGF-β may reflect impaired anti-inflammatory and vascular homeostatic processes (Fig. 3h). Functionally, these proteins were also associated with cognitive function. Regression analysis showed that MoCA scores were positively associated with IL-4 levels (Fig. 3i), while higher FAS scores correlated with increased CD33 and decreased PDGF-β, and elevated TNF levels were associated with higher PASC scores.

A multivariable logistic model using the least absolute shrinkage and selection operator (LASSO)⁴⁰ method on 54 DEPs (44 Cog-PASC-related and 10 immune/vascular regulatory proteins) identified CASQ2, ARHGAP1, and TREM2 as key predictors (Fig. 3j, k), achieving an area under the curve of 0.889 in the receiver operating characteristic analysis. These findings suggest that these proteins alone effectively identify individuals at risk for Cog-PASC, underscoring their potential pathogenic relevance.

Taken together, the Cog-PASC group exhibited distinct DEPs related to oxidative stress and synaptic dysfunction, along with alterations in immune and vascular regulatory proteins. Many of these proteins were significantly associated with cognitive performance, and several (CASQ2, ARHGAP1, and TREM2) emerged as independent predictors of Cog-PASC in multivariable modeling. These findings suggest that pathways involving synaptic regulation, oxidative stress responses, and immune and vascular interactions may contribute to the pathogenesis of Cog-PASC and could help identify individuals at risk.

Brain alterations in Cog-PASC replicated in an independent cohort

We used the validation cohort (Fig. 1) as an independent dataset to replicate the important findings from the exploration cohort. A total of 109 patients were included: 16 with Cog-PASC, 47 with Other-PASC, and 46 with NS-PASC (Table 5). The cohort comprised 72 females (66.1%). Participants were enrolled a median of 16.0 months after infection [IQR 10.0–22.0], and MRI was performed at a median of 18.0 months [IQR 11.0–23.0] post-infection, both significantly longer than in the exploration cohort (p < 0.001 for both). As in the exploration cohort, most participants were infected during the Omicron-dominant period.

The Cog-PASC group tended to be older, less educated, and less frequently female; these variables were adjusted for in subsequent analyses. This group showed more severe cognitive impairment, whereas the Other-PASC group had higher fatigue scores. Overall, PASC scores and symptom prevalence were comparable between the Cog-PASC and Other-PASC groups (Table 5 and Supplementary Fig. 6a, b).

Blood biomarker analysis using Simoa technology replicated the finding of elevated brain damage-associated proteins in the Cog-PASC group; both GFAP and NfL levels were significantly higher than in the other groups, suggesting more pronounced neuronal and glial damage (Fig. 4a, b). MRI analysis confirmed widespread cortical thinning in the Cog-PASC group, most notably in the frontal, cingulate, insular, and parietal cortices (Table 6 and Fig. 4c). Cingulate cortical thickness was positively associated with MoCA scores, independent of demographic covariates (Fig. 4d). Volumetric analyses revealed significant volume reductions in the whole cortex and parietal cortex.

The Cog-PASC group also showed increased paramagnetic susceptibility in the occipital cortex, caudate, putamen, and hippocampus, consistent with the exploration cohort (Fig. 4e), although susceptibility values were not associated with MoCA scores (Fig. 4f). Additional structural measures showed a significantly larger choroid plexus volume and a reduced DTI-ALPS index in the Cog-PASC group (Fig.4g), although no significant associations were observed between choroid plexus volume or DTI-ALPS index and MoCA scores (Fig. 4h). In DTI-based TBSS analysis, the Cog-PASC group demonstrated elevated MD and RD alongside reduced FA across bilateral white matter regions compared with the NS-PASC and Other-PASC groups (P_FWE < 0.05), a pattern suggestive of microstructural white matter injury and potential demyelination²⁰ (Table 6).

Taken together, the validation cohort reproduced key findings from the exploration cohort, confirming that individuals with a history of mild COVID-19 who present with cognitive symptoms exhibit structural brain changes suggestive of neurodegeneration, including cortical thinning, volume loss, iron accumulation, CP enlargement, and microstructural white matter injury.

This study identified distinct structural and neurodegenerative changes in the brains of individuals with Cog-PASC that were not evident in other-PASC groups, most of whom had not been hospitalized during the acute phase, that were not observed in other-PASC groups. These results were independently replicated in a separate validation cohort, suggesting that Cog-PASC may involve unique neuropathological processes and highlighting the need for targeted monitoring strategies in affected individuals.

Although brain injury has been reported in moderate-to-severe cases of acute COVID-19⁴¹, whether this damage persists and leads to chronic neurodegeneration is unclear. Recent biomarker studies, primarily focused on post-hospitalized individuals, have suggested increased brain injury in the chronic phase^10,42. However, other studies, particularly those involving non-hospitalized individuals, did not confirm persistent brain damage^43,44, indicating limited ongoing injury. In this study, which included a larger cohort of predominantly non-hospitalized individuals^43,44, we found evidence of chronic brain injury and neurodegenerative changes in the Cog-PASC group across both the exploration and validation cohorts. Specifically, elevated GFAP and/or NfL levels were observed without corresponding alterations in β-amyloid or phosphorylated tau proteins. This is notable, as SARS-CoV-2 infection has been linked to AD pathogenesis, with viral proteins potentially forming structures that may drive amyloid formation, and in vitro data showing increased tau phosphorylation and aggregation due to SARS-CoV-2⁴⁵. The brain injury observed in this population, approximately one year post-infection, may represent a distinct process from classical AD, where mild cognitive impairment typically involves β-amyloid and tau abnormalities⁴⁶. These results warrant further neuropathologic investigation.

Cortical thickness was significantly reduced in the cingulate cortex, parahippocampal gyrus, and insular cortex of the Cog-PASC group compared with the Other-PASC group in the exploration cohort, with only modest differences relative to the NS-PASC group. In the covariate-matched cohort, cingulate cortical thickness remained significantly lower in the Cog-PASC. These findings were confirmed in the validation cohort, where the Cog-PASC group showed the greatest cortical thinning in the cingulate and insular cortices, as well as in the frontal and parietal cortices. The affected regions are consistent with those reported in previous post-COVID brain studies^5,7,8. Meanwhile, brain volume analyses in the exploration cohort revealed no significant differences among PASC groups, whereas in the validation cohort, the Cog-PASC group showed volume reductions in the whole cortex and parietal cortex. These discrepant results suggest that brain volume may be less sensitive to detecting post-COVID structural changes. This aligns with previous reports indicating that cortical thickness is less influenced by intracranial volume and race and is more sensitive to pathological aging than volumetric measures^47,48,49, thereby providing a more reliable marker of subtle neuroanatomical change. The greater disparity in cortical thickness between the Cog-PASC and Other-PASC groups, compared with the Cog-PASC and NS-PASC groups, warrants further discussion. Cortical thinning in the Cog-PASC group could reflect a reduced brain functional reserve⁵, which indicates vulnerability to cognitive impairment⁸, rather than being a result of significant brain damage. Another possibility is cortical enlargement in the Other-PASC group, possibly due to neurogenesis or compensatory mechanisms, as reported particularly in mild COVID cases^50,51. Neurogenesis, potentially supported by enhanced CP function, may increase gray matter⁵², while compensatory mechanisms could cause cortical hypertrophy⁵³. Despite modest cortical thinning, the Cog-PASC group showed elevated brain damage-associated proteins, decreased oxidative stress-related proteins, and increased inflammatory markers, which are molecular signatures associated with gray matter loss and potential long-term structural vulnerability.

Using the χ-separation method to isolate paramagnetic iron susceptibility, we observed increased cortical iron deposition in the Cog-PASC group within the exploration cohort, most notably in the cingulate, occipital cortex, and hippocampus, compared with the NS-PASC and Other-PASC groups. In the covariate-matched cohort, iron accumulation was most prominent in the cingulate. Analysis of the validation cohort further confirmed elevated iron levels, particularly in the occipital cortex, hippocampus, caudate, and putamen. Iron accumulation may contribute to neurodegeneration through two key mechanisms^11,54: (1) generating reactive oxygen species that induce oxidative damage and ferroptosis^11,54,55 and (2) promoting the aggregation of neurodegenerative proteins⁵⁴. Excess iron may also aggravate BBB disruption by inducing oxidative stress and neuroinflammation, leading to microglial activation and the release of cytokines and matrix metalloproteinases that degrade tight junctions and basement membranes⁵⁶. Chronic iron overload further promotes endothelial ferroptosis and vascular remodeling⁵⁶, potentially amplifying vascular injury initiated by SARS-CoV-2 binding to ACE2 receptors on brain microvascular endothelial cells⁵⁷.

Notably, the covariate-matched cohort, which minimizes confounding by age, sex, and education, revealed a convergence of structural and metabolic alterations in the cingulate cortex of Cog-PASC individuals, including both significant cortical thinning and iron deposition. This dual burden suggests that the cingulate may serve as a key substrate for persistent cognitive impairment in long COVID. The cingulate cortex is essential for attentional control, emotional regulation, and cognitive integration, and is particularly sensitive to neuroinflammatory and oxidative stress–mediated injury^58,59. Iron accumulation may further exacerbate regional vulnerability by promoting microglial activation and synaptic dysfunction, while cortical thinning may reflect neuronal or glial loss. These findings are consistent with prior reports of anterior cingulate thinning in post-COVID patients⁵ and with broader literature implicating the cingulate in cognitive impairment across neuropsychiatric conditions⁶⁰.

The CP was larger in the Cog-PASC group compared to the Other-PASC group, suggesting that structural alterations in the CP may contribute to the cognitive manifestations of PASC. As a multifunctional structure, the CP forms the blood–CSF barrier, produces CSF, supplies nutrients, and supports waste clearance and immune surveillance¹⁴. Structural changes in the CP, such as epithelial flattening and vessel wall thickening, have been observed in aging as well as in patients with AD⁶¹. MRI studies have also reported increased CP volume in individuals with mild cognitive impairment, AD, and PD^15,16, suggesting that CP enlargement may serve as a biomarker of CP dysfunction⁶².

DTI-ALPS and other diffusivity measures (MD, RD, and FA) showed no statistically significant differences in the exploration cohort, although the direction of change in the Cog-PASC group (higher MD and RD, lower FA) was consistent with our hypothesis. In the validation cohort, however, the DTI-ALPS index was significantly reduced, while MD and RD were increased and FA decreased, indicating more pronounced microstructural alterations. A lower DTI-ALPS index suggests impaired glymphatic function and reduced interstitial fluid clearance, whereas increased MD and RD, together with decreased FA, are compatible with white matter injury involving demyelination or axonal degeneration. Similar reductions have been observed in COVID-19 patients three months post-infection⁶³. These findings support the view that long COVID may involve both functional glymphatic disruption and structural compromise of white matter integrity, warranting further investigation in larger long COVID cohorts.

Blood proteomic analyses revealed significant alterations in proteins involved in oxidative stress and synaptic function, which are both associated with neurodegeneration. A multivariable model identified TREM2, ARHGAP1, and CASQ2 as key DEPs independently associated with cognitive impairment. TREM2, a microglial receptor responsible for clearing pathological proteins and promoting neuroprotection⁶⁴, is markedly upregulated in severe COVID-19, particularly in T cells and mononuclear cells, where it enhances immune activation and proinflammatory signaling via interactions with the SARS-CoV-2 membrane protein⁶⁵. This upregulation may contribute to persistent neuroinflammation and immune dysregulation, potentially accelerating neurodegenerative processes⁶⁵. ARHGAP1 regulates Rho GTPase activity, influencing cytoskeletal remodeling and cell polarity⁶⁶. Although its direct role in neurodegeneration is unclear, dysregulation of Rho/ROCK signaling may affect microglial function and synaptic pruning⁶⁷. CASQ2 is a calcium-binding protein that regulates calcium storage, primarily in cardiac muscle⁶⁸, and is the predominant isoform in the hippocampus, where it supports neuronal calcium balance and synaptic plasticity⁶⁸. Disruption of CASQ2 may impair calcium homeostasis and contribute to neuroinflammation and neurodegeneration.

The convergence of structural MRI findings and proteomic alterations in the Cog-PASC group provides insight into potential pathophysiologic mechanisms. Cortical thinning, particularly in the cingulate and insular cortices, may be linked to the observed downregulation of synaptic regulatory proteins, reflecting impaired synaptic maintenance and neuronal connectivity. Similarly, increased cortical and subcortical iron deposition could exacerbate oxidative stress, consistent with the reduced abundance of antioxidant enzymes such as SOD1 and CCS detected in the proteomic analysis. CP enlargement, observed on MRI, may also contribute to impaired waste clearance and chronic neuroinflammation, potentially sustaining the upregulation of inflammatory proteins including TREM2, CXCL10, and VCAM-1. Together, these findings suggest that structural brain alterations and molecular signatures are part of a converging cascade involving synaptic dysfunction, oxidative stress, and immune dysregulation, which may underlie the persistent cognitive deficits observed in Cog-PASC.

Long-COVID symptoms, aside from cognitive dysfunction, did not significantly differ between Other-PASC and Cog-PASC groups. Notably, FAS and PASC scores, unlike cognitive symptom scores, showed no association with blood markers of brain injury or structural changes, but instead correlated with several immune and vascular regulatory proteins. Specifically, higher FAS scores were linked to increased CD33 and decreased PDGF-β, while higher PASC scores were associated with elevated TNF levels. Elevated CD33 expression in severe COVID-19 has been associated with immunosuppressive myeloid responses, which may indirectly suppress T-cell activity⁶⁹, impair antiviral defense, and contribute to chronic inflammation and fatigue. Reduced PDGF-β levels can impair vascular repair by disrupting pericyte recruitment and endothelial stability, thereby worsening hypoxia and tissue stress⁷⁰. Elevated TNF levels in PASC may drive persistent inflammation through sustained activation of TNF-producing T cells and macrophages, leading to tissue damage and fatigue^71,72. TNF also synergizes with IL-1β and IL-6 to amplify NF-κB signaling and exacerbate endothelial dysfunction^71,72. It is worth noting that our previous study demonstrated that patients with the PASC symptom cluster, including fatigue and post-exertional malaise, had different blood cytokine profiles compared to those with the PASC symptom cluster, including cognitive dysfunction (KJIM in major revision). The course of long-COVID symptoms, particularly cognitive impairment, may be influenced by patient-specific factors such as brain functional reserve and CP volume, leading to differentiation into Cog-PASC and Other-PASC groups. Hence, our findings suggest that cognitive symptoms may have different pathophysiologic mechanisms from other-PASC symptoms. Further studies are needed for these distinct Cog-PASC patients to elucidate biomarkers for diagnosis and potential therapeutic targets.

Our study has several limitations. First, it employed a cross-sectional design, which identifies associations but not cause-and-effect relationships. As this is a preliminary report from a longitudinal study with follow-up MRI findings, future reports will provide more details on the clinical courses and long-term consequences of MRI findings and blood proteomic signatures. Second, this study lacks stratification by case severity, as the majority (~98%) of participants were non-hospitalized patients with only mild acute symptoms. Therefore, our results may not apply to patients who experienced more severe COVID-19 symptoms. Third, baseline characteristics such as age, sex, and education were not evenly distributed across groups. Although adjustments were made for these variables, results should be interpreted with caution. Fourth, the exploration and validation cohorts were not fully comparable, with differences such as enrollment and MRI timing. This may limit comparability, and the relatively small sample size further increases susceptibility to random variation, potentially contributing to discrepancies between cohorts, particularly in volumetric and DTI analyses. Nonetheless, the consistent direction of changes across cohorts supports the biological relevance of the findings. Finally, this study was conducted at a single tertiary medical center in Korea within a single ethnic group, limiting the generalizability of the findings.

In conclusion, our study suggests distinct brain structural alterations and ongoing neurodegenerative processes in individuals with cognitive dysfunction, which are not prominent in those with other symptoms among PASC. Hence, the pathogenesis underlying cognitive dysfunction in long COVID differs from non-cognitive symptoms, while its long-term course and clinical significance should be confirmed in future longitudinal studies.

This is a preliminary report from the exploration cohort of the LOCOMOTIVE study (LOng Covid Outcome MOnitoring with Translational InVEstigations of blood and neuroimaging), an ongoing prospective, longitudinal cohort study investigating neurological symptoms and brain structural changes associated with long COVID. The study evaluates longitudinal changes in symptoms, brain MRI findings, and blood profiles. Brain MRI scans are conducted at baseline and 1 year to assess structural changes, including gray matter alterations. Blood samples are collected at baseline, 6 months, and 1 year, with symptom evaluations every 3 months.

Registered with the Clinical Research Information Service (cris.nih.go.kr, KCT0007848), participants for the exploration cohort were recruited between November 2022 and October 2023 from a tertiary medical center (Asan Medical Center, Seoul, South Korea), followed by validation cohort recruitment from November 2023 to April 2025. This preliminary analysis focused on potential brain damage, structural alterations, and neurodegenerative processes following SARS-CoV-2 infection.

Healthy individuals with no prior cognitive complaints or impairments in daily living activities before contracting SARS-CoV-2 were eligible for enrollment; however, their data were not included in this study. Exclusion criteria included cognitive complaints before COVID-19, acute illnesses within the past two months (e.g., myocardial infarction, unstable angina, pneumonia), active cancer⁷³, autoimmune diseases requiring immunosuppressants, neurodegenerative conditions (e.g., Alzheimer’s disease, Parkinson’s disease), and psychiatric disorders.

Participants underwent blood sampling and assessments using standardized tools, including the Montreal Cognitive Assessment (MoCA) for cognitive function⁷⁴, the Fatigue Assessment Scale (FAS) for fatigue⁷⁵, and detailed questionnaires to evaluate post-acute sequelae of COVID-19 (PASC) symptoms^76,77. The PASC symptoms assessed included post-exertional malaise, fatigue, brain fog, dizziness, gastrointestinal symptoms, palpitations, changes in sexual desire or capacity, loss or change in smell or taste, thirst, chronic cough, chest pain, and abnormal movements¹. Symptom severity during the acute phase of COVID-19 was determined using the World Health Organization (WHO) ordinal scale⁷⁸.

Sex was recorded as assigned at birth (ascertained from medical records at enrollment). Although sex was not a primary variable of interest in the study design, baseline differences between groups were observed. Therefore, sex was included as a covariate in statistical analyses to adjust for potential confounding effects.

All participants provided written informed consent in accordance with the Declaration of Helsinki. The study protocol was approved by the Institutional Review Board of Asan Medical Center (IRB No. 2022–1378). Transportation and examination costs for study visits were reimbursed, but no additional financial or other benefits were provided.

Classification of the participants

Participants with SARS-CoV-2 infection were categorized into three groups based on cognitive impairment and other post-acute sequelae of COVID-19 (PASC). The Cog-PASC group included individuals with cognitive impairment (MoCA score <26)⁷⁴. The Other-PASC group comprised those without cognitive impairment but with significant fatigue (FAS score ≥22)⁷⁵ and/or other-PASC symptoms (PASC score ≥12)^76,77. Patients who did not meet the criteria for either of these groups were classified into the no significant PASC (NS-PASC) group.

Matching cohort

Due to significant differences in age, sex, and education level among the PASC groups, we constructed a covariate-matched cohort to minimize potential demographic confounding. Participants were matched in a 1:1 ratio using nearest-neighbor matching without replacement, based on propensity scores estimated from a logistic regression model that included age, sex, and education level as covariates. A caliper width of 1.5 standard deviations of the logit of the propensity score was applied for the Non-PASC versus Cog-PASC comparison, and 2.0 standard deviations for the Other-PASC versus Cog-PASC comparison. Matching balance was evaluated using standardized mean differences (SMDs) for each covariate, with an SMD below 0.1 considered indicative of adequate balance. After matching, group differences were assessed using independent t-tests for continuous variables (age, education level) and Fisher’s exact test for the categorical variable (sex).

Blood sampling and protein analysis

Blood samples were collected and stored at −80 °C following standardized procedures⁷⁹, then analyzed using advanced platforms. Brain damage markers, including GFAP, NFL, p-Tau181, and the Aβ40/42 ratio, were measured using the Simoa HD-X platform. Assay results were reported in pg/mL, with data normalized according to the manufacturer’s protocol. Proteomic analysis utilized the Olink Explore 3,072 platform, which measures proteins linked to inflammation, neurology, oncology, and cardiometabolic pathways using Proximity Extension Assay (PEA) technology. Protein concentrations were transformed into NPX values on a log₂ scale, where a 1 NPX increase equates to a protein concentration doubling.

Differentially expressed protein analysis

Multivariate linear regression, adjusted for age and sex, identified differentially expressed proteins (DEPs) with adjusted p-values < 0.05 using the Benjamini–Hochberg method to control the false discovery rate (FDR). DEPs were visualized using volcano plots, with additional box and principal component analysis (PCA) plots generated in R (ggplot2, version 4.4.1).

Gene set enrichment and pathway analysis

Enrichment analysis of DEPs used the Human Protein Atlas (HPA) dataset³³ and Gene Ontology (GO) terms, including biological processes, molecular functions, and cellular components³⁴. Bubble plots visualized functional associations in R. Pathway analysis integrated data with the Kyoto Encyclopedia of Genes and Genomes (KEGG) to map interactions and pathways³⁵. Protein–protein interaction (PPI) networks provided an intuitive view of GO terms’ significance and prevalence.

Protein–protein interaction network analysis

A PPI network was created using the STRING database (minimum interaction score: 0.4)³⁶, with nodes representing proteins and edges indicating functional associations. Stronger interactions were shown with thicker edges. Key clusters were identified through K-means clustering, with centrality assessed by node degree—higher degrees indicated greater connectivity and biological significance.

MRI acquisition

MRI data were collected using 3.0-Tesla scanners (Ingenia CX, Philips Healthcare) with a 32-channel head coil, following protocols tailored for neurodegenerative diseases. Sequences included 3D T1-WI for cortical parcellation, DTI for ALPS index calculation, and multiecho gradient echo (GRE) for paramagnetic susceptibility maps. Detailed acquisition parameters are in Supplementary Table 2.

Cortical parcellation using FreeSurfer

Cortical thickness and its volumes were measured on 3D T1-WI using FreeSurfer (v.7.3)^80,81. Processing steps included removing non-brain tissue, Talairach transformation, intensity normalization, and segmentation of gray (GM) and white matter (WM). The cortical surface was parcellated into regions of interest (ROIs) using the Desikan-Killiany Atlas⁸², then grouped into lobar regions such as the frontal, parietal, temporal, occipital, insular, and cingulate cortices⁸³. Additionally, following a previous report on greater reductions of cortical thickness in the orbitofrontal cortex, cingulate cortex, and parahippocampal gyrus among patients with COVID infection⁵, we included these regions a priori.

Paramagnetic susceptibility maps

Multiecho GRE images were post-processed to generate paramagnetic susceptibility maps⁸⁴. Phase maps were unwrapped using a Laplacian algorithm⁸⁵, and background field removal isolated tissue-specific fields. R2* maps were derived by fitting a mono-exponential model to the GRE signal magnitude. These were applied to susceptibility source separation using the χ-separation toolbox, producing paramagnetic susceptibility maps³². Cortical and deep GM regions were segmented using SynthSeg⁸⁴, a tool optimized for diverse image contrasts and resolutions. Segmentation masks were directly overlaid onto the susceptibility maps to obtain susceptibility values without requiring co-registration.

Choroid plexus segmentation

Choroid plexus volumes (CPVs) were segmented from FreeSurfer-derived 3D T1-WI masks of the lateral and inferior lateral ventricles. Using a Gaussian mixture model (GMM), voxels were grouped by intensity, distinguishing cerebrospinal fluid (CSF) from choroid plexus voxels⁸⁶. The GMM method outperforms FreeSurfer’s segmentation^87,88. All automated masks were reviewed and corrected by an experienced neuroradiologist. CPV was normalized by dividing by total intracranial volume (TIV), yielding the CPV fraction.

DTI-ALPS processing

DTI-ALPS preprocessing followed established methods^87,88,89, with ALPS indices calculated using a MATLAB-based in-house program (DTI-ALPS Analyzer, MATLAB 9.6, version 2019b, the Mathworks, Inc). The detailed calculation steps are provided in Supplementary Methods. Diffusion coefficient maps (Dx, Dy, Dz) were generated, and fractional anisotropy (FA) maps were used to define ROIs on projection, association, and subcortical white matter fibers. Subject-specific 5-mm spherical ROIs were manually drawn on the slice containing the uppermost portion of the left lateral ventricle using color-coded FA maps, where three ROIs were defined according to their directions along the projection, association, and subcortical fibers (Supplementary Fig. 7). A blinded radiologist with 11 years of experience performed the delineation. After a ~30-week washout period, the ROIs were redrawn to assess intraobserver reliability, yielding an intraclass correlation coefficient (ICC) of 0.603 (95% confidence interval, 0.49–0.69; P < 0.001), comparable with previous reports⁹⁰.

The ALPS index, calculated as the mean diffusion in projection and association fibers divided by the mean in orthogonal fibers [mean(Dxassoc, Dxproj)/mean(Dzassoc/Dyproj)], reflects the water molecules’ diffusivity along the radial direction at the lateral ventricular body level. A lower ALPS index suggests impaired water diffusion along perivascular spaces and potential dysfunction in waste clearance⁸⁹. Furthermore, voxel-wise comparisons of mean diffusivity (MD), radial diffusivity (RD), and FA among the three groups were performed using Tract-Based Spatial Statistics (TBSS), a widely used method for spatial statistical analysis based on the white matter skeleton⁹¹. TBSS enables precise detection of microstructural alterations in white matter and has become a widely adopted approach for investigating white matter fiber tracts^20,92,93. The analysis included age, sex, and education as covariates for adjustment. Multiple comparison correction was performed using the family-wise error (FWE) method, with statistical maps thresholded at p < 0.05 (two-sided).

Statistical analysis

All statistical analyses were performed using R version 4.1.1. (R Foundation for Statistical Computing, Vienna, Austria). Analyses were conducted separately in the exploration cohort, the covariate-matched cohort constructed based on age, sex, and education level, and the independent validation cohort. Descriptive statistics were used to summarize demographic and clinical variables, including medians and interquartile ranges (IQRs) for continuous variables and frequencies and percentages for categorical variables. Group comparisons were conducted using the Kruskal–Wallis test with Dunn’s correction for multiple comparisons. Additionally, analysis of covariance (Two-sided ANCOVA) was performed to compare biomarker levels across the three groups, adjusting for age, sex, and education as covariates. Group differences were assessed by comparing adjusted means within the model, with Tukey’s Honest Significant Difference (HSD) test used for post hoc multiple comparisons. For MRI parameter comparisons across the three groups, false discovery rate (FDR) correction was applied to control for multiple comparisons. Intraclass correlation coefficients (ICC) were calculated using a two-way consistency model.

To identify differentially expressed proteins (DEPs) relevant to Cog-PASC, volcano plot analysis was conducted based on multivariate linear regression, with p-values corrected for multiple comparisons using the Benjamini–Hochberg method to control the FDR. Pearson correlation matrices were constructed to assess the linear relationships between biomarkers, clinical variables, and MRI parameters. The multivariable logistic regression model for cognitive impairment prediction was constructed using the least absolute shrinkage and selection operator (LASSO) method to optimize variable selection and prevent overfitting. Model performance was evaluated using multiple metrics. The discriminative ability of the model was assessed using the area under the Receiver Operating Characteristic curve (AUC-ROC), while diagnostic accuracy was measured through sensitivity and specificity calculations. The overall accuracy of the model and its corresponding confidence intervals were calculated to determine model stability. Additional model validation was conducted using McFadden’s R² and Kappa statistics, while model calibration was assessed using McNemar’s test.

Reporting summary

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