Disrupted Structure–Function Integration in Systemic Lupus Erythematosus and Its Impact on Cognitive Flexibility

ABSTRACT

Systemic lupus erythematosus (SLE) is a chronic autoimmune disease, with cognitive dysfunction being one of its most common neuropsychiatric manifestations. Cognitive flexibility relies on the integration of brain structure and function, with white matter networks providing anatomical constraints for functional dynamics. Reduced cognitive flexibility is frequently observed in SLE, but the underlying structure–function integration changes remain poorly understood. This study investigated whether brain structure–function integration is altered in SLE and how it links to cognitive flexibility. We examined 22 SLE patients without clinically overt neuropsychiatric manifestation (age: 34.99 ± 10.67; 18 females) and 60 healthy controls (HCs) (age: 28.43 ± 8.56; 29 females). Using diffusion MRI and task‐based fMRI acquired during the Montreal Card Sorting Test (MCST), a cognitive flexibility task, we derived brain structural–functional alignment and liberality, which quantify the extent to which brain functional signals are either coupled with or deviate from the underlying anatomical network. We found SLE patients exhibited globally higher liberality and lower alignment compared to HCs, and this was driven by the disrupted structure–function integration in the executive control network (ECN). The ECN comprises three subnetworks: ECN‐A and ECN‐B comprise key lateral fronto‐parietal executive control areas, while ECN‐C is anatomically closer to the default mode network. Further analyses revealed that SLE had higher liberality in ECN‐A and ECN‐B regions, alongside lower alignment in ECN‐A, while ECN‐C did not show these alterations. Importantly, increased liberality and decreased alignment in the ECN regions were associated with poorer cognitive flexibility (MCST performance) in SLE participants. This association was also observed across all participants. In SLE individuals specifically, liberality and alignment in the fronto‐parietal ECN were further linked to clinical variables, including serum albumin and corticosteroid dosage. Additionally, the liberality and alignment in the ECN and its subnetworks were associated with cognitive performance outside the scanner (measured by Automated Neuropsychological Assessment Metrics) across all participants. Our findings suggest that aberrant structure–function integration, particularly within the fronto‐parietal ECN, impacts cognitive flexibility and may contribute to the development of cognitive impairment in SLE.

This study compared the brain structure–function integration during a cognitive flexibility task between SLE patients without clinically overt neuropsychiatric manifestations and healthy controls. Our findings suggest that disrupted structure–function integration, particularly within the frontoparietal executive control network (ECN) regions, influences cognitive flexibility and might contribute to the cognitive dysfunction in SLE.

1. Introduction

Systemic lupus erythematosus (SLE) is an immune‐complex mediated, multi‐systemic autoimmune disease with a broad spectrum of clinical manifestations (Siegel and Sammaritano 2024; West 1996). Involvement of the central nervous system (CNS) that manifests as various neurological and psychiatric symptoms is reported in over 50% of SLE patients during their disease course (El‐Shafey et al. 2012; Schwartz et al. 2019). Cognitive dysfunction is one of the commonest neuropsychiatric syndromes in SLE, which is defined as significant deficits in any or all of the following cognitive functions in the American College of Rheumatology (ACR) nomenclature that includes executive skills, complex attention, memory, visual–spatial processing, language, and psychomotor speed (Liang et al. 1999; Seet et al. 2021). Although factors such as autoantibodies, cytokine‐mediated neuronal dysfunction (Barraclough et al. 2019; Ceccarelli et al. 2018; El‐Shafey et al. 2012; Zabala et al. 2018), glucocorticoid exposure (Seet et al. 2021), and cerebrovascular disease (Barraclough et al. 2019) have been implicated in the development of cognitive dysfunction in SLE, the underlying pathophysiology and neural substrates, particularly system‐level brain network organization, remain poorly understood (El‐Shafey et al. 2012; Leslie and Crowe 2018).

Impairment in executive function is a commonly reported and functionally significant component of cognitive dysfunction in SLE (Maciel et al. 2016; Mak et al. 2012; Seet et al. 2021). Executive function encompasses a set of higher‐order cognitive processes, such as working memory, shifting, and inhibition, that support goal‐directed behavior (Best and Miller 2010). Among these, cognitive flexibility is a core component, referring to the ability to adapt to changing circumstances, shift between mental sets or concepts, and modify strategies to achieve behavioral goals (Dajani and Uddin 2015; Hohl and Dolcos 2024). For example, it allows individuals to adjust problem‐solving strategies when initial approaches are unsuccessful, such as shifting from rote memorization to logical reasoning when faced with a complex task. Cognitive flexibility is typically measured using set‐shifting or task‐switching behavioral paradigms, in which participants must switch between tasks with different instructions given some stimuli (Dajani and Uddin 2015). The successful implementation of cognitive flexibility relies on the coordinated functioning of multiple executive function processes, including salience detection, attention, working memory, inhibition, and switching (Dajani and Uddin 2015; Uddin 2021). In dynamic environments, individuals must detect relevant changes, inhibit previously appropriate but now ineffective responses, and reconfigure mental strategies in real time to flexibly switch toward new goals. The cognitive switching process is considered the most complex form of cognitive flexibility (Crone et al. 2006). In individuals with SLE, approximately 39% exhibit impairments in executive function and complex attention (Monahan et al. 2021). Compared to matched healthy controls (HCs), SLE patients show reduced cognitive flexibility, as evidenced by lower performance on the Dimensional Change Card Sort Test (Plantinga et al. 2018, 2024). These cognitive impairments have been linked to greater organ damage and specific serum analyte profiles (Muñoz‐Grajales et al. 2024; Raghunath et al. 2023), and also are associated with diminished performance in activities of daily living and reduced participation in life roles (Barraclough et al. 2019).

Functional neuroimaging studies of cognitive flexibility have identified a distributed network of frontoparietal regions involved in flexible switching, including high‐level cortical association areas (ventrolateral prefrontal cortex, dorsolateral prefrontal cortex, anterior cingulate, anterior insula), the pre‐motor cortex, the inferior and superior parietal cortices, the inferior temporal cortex, the occipital cortex, and subcortical structures such as the caudate and thalamus (Dajani and Uddin 2015; Leber et al. 2008; Uddin 2021). These regions interact dynamically to support the ability to shift attention, update goals, and implement new behavioral strategies, which are core aspects of cognitive flexibility. This flexible control system broadly aligns with what is often referred to as the executive control network (ECN), which can be further subdivided into distinct subnetworks based on large‐scale functional parcellation schemes such as that proposed by Thomas Yeo et al. (2011). In the 17‐network solution, the ECN includes ECN‐A and ECN‐B, which primarily cover lateral frontoparietal regions (e.g., dlPFC and inferior parietal lobule) and are engaged in externally focused, goal‐directed cognition. A third subnetwork, ECN‐C, is positioned closer to default mode regions and may subserve more internal or self‐referential aspects of cognitive control. These ECN subnetworks are thought to play differential roles in supporting cognitive flexibility, and may be selectively disrupted in SLE patients with cognitive dysfunction (Kozora et al. 2016).

The importance of the brain's structural constraints on functional communication in supporting executive function and cognitive flexibility is increasingly evident (Baum et al. 2020; Suárez et al. 2020). Particularly, a previous study reported that cognitive processes requiring flexibility are supported by multimodal integration of brain network anatomy and functional dynamics (Medaglia et al. 2018). Specifically, the study decomposed the functional blood‐oxygenation‐level‐dependent (BOLD) signal during a cognitive switching task into components that either aligned with or deviated from the underlying anatomical network, defined by individual‐level white matter (WM) connectivity derived from diffusion MRI. They found that whole‐brain functional liberality, reflecting the degree of deviation from the anatomical network, was positively associated with task switching cost in healthy young adults. This suggests that liberality captures functionally meaningful neural activity that is less anatomically constrained and may reflect increased cognitive demands that require flexibility. This finding is further supported by recent studies that employed similar methods, defining structural‐functional decoupling as the ratio of the norms of liberal to aligned signal components (Dong et al. 2024; Griffa et al. 2022; Preti and Van De Ville 2019). These studies demonstrate that this decoupling is behaviorally meaningful: brain regions with strong structural–functional coupling (i.e., functionally more aligned with anatomy) are typically involved in sensorimotor and perceptual functions, such as multisensory integration, visual perception, motor and eye movements, and auditory processing, that prioritize efficiency. In contrast, regions with greater structural–functional decoupling (i.e., functionally more liberal from anatomy) are more engaged in higher‐order processes such as reward, emotion, affective regulation, social cognition, semantic processing, memory, and cognitive control, which require greater flexibility.

An accumulating body of structural and functional MRI research has identified global and regional brain abnormalities associated with cognitive dysfunction in patients with SLE (Cabrera et al. 2024). Meta‐analyses of resting‐state fMRI studies have consistently demonstrated that functional alterations in SLE are predominantly localized to the default mode network (DMN) and the limbic system (Wang et al. 2024). Notably, a recent study reported that abnormalities in whole‐brain resting‐state functional connectivity accounted for a substantial portion of the association between extensive blood–brain barrier leakage and cognitive impairment in SLE (Hanly et al. 2023). Complementing these functional findings, recent diffusion‐weighted MRI studies revealed that disruptions in WM microstructure and structural network integrity may contribute to the development of severe neuropsychiatric symptoms in SLE (Bai et al. 2025; Hu et al. 2024). Compromised WM microstructural integrity has also been observed in the early stages of SLE—even in the absence of overt neuropsychiatric symptoms (Silvagni et al. 2021). In addition, increased volumes of WM hyperintensities have been linked to poorer cognitive performance in SLE (Monahan et al. 2021). Our previous work also found elevated extracellular free water in WM, which was significantly associated with deficits in sustained attention (Qian et al. 2022). Despite these findings, it remains largely unexplored whether and how the structural constraints imposed by the WM network on functional brain activity are altered in individuals with SLE—particularly during cognitively demanding tasks that require cognitive flexibility. Addressing this gap could offer critical insights into the neurobiological mechanisms underlying cognitive dysfunction in SLE.

In this study, we aimed to compare brain structural–functional integration (alignment and liberality) between SLE and HC using diffusion MRI and fMRI data while they were performing the Montreal card sorting test (MCST) (Au et al. 2012; Monchi et al. 2006). The MCST measures cognitive flexibility by examining how adaptive participants were toward an implicit and shifting classification rule (Stad et al. 2019). Based on previous findings (Medaglia et al. 2018), we hypothesized that functional signals would be more aligned (and/or less liberal) with the anatomical network, specifically the ECN, during cognitive switching in SLE patients. We further hypothesized that the brain structural–functional integration would be associated with cognitive flexibility performance as well as out‐of‐scanner neurocognitive performance in SLE. Given the multifactorial nature of cognitive dysfunction in SLE, we also sought to examine the relationship between structural‐functional integration and clinical or immunological indicators of disease activity. We hypothesized that greater systemic disease burden or elevated disease activity would be linked to more pronounced disruptions in structural–functional integration relevant to cognitive flexibility.

2. Methods

2.1. Participants

We recruited 22 SLE patients without clinically overt neuropsychiatric manifestation from the Lupus Clinic of the National University Hospital (NUH) Medical Centre, Singapore. All patients met either the 1997 ACR criteria (Hochberg 1997; Tan et al. 1982) or the 2012 Systemic Lupus International Collaborating Clinics (SLICC) classification criteria for SLE (Petri et al. 2012). A group of 60 HCs was recruited from the local community. Exclusion criteria for both SLE patients and HCs included: (1) current use of anticoagulant or antiplatelet medications; (2) inability to understand the study procedures; (3) history of neurological disorders, including cerebrovascular disease or seizures; (4) symptoms of severe depression and/or anxiety, assessed using the Hamilton Depression Rating Scale (HAM‐D) and the Hamilton Anxiety Rating Scale (HAM‐A) (Mak et al. 2011; Zigmond and Snaith 1983). Additional exclusion criteria for SLE patients included: (1) recent augmentation of immunosuppressive therapy, including glucocorticoid use within the past 2 weeks; (2) diagnosis of autoimmune overlap syndromes. All participants provided written informed consent prior to participation. The study was approved by the local ethics review board (DSRB reference: 2014/00766). MRI scans were conducted at the Centre for Translational Magnetic Resonance Research, National University of Singapore. Of the 82 participants initially recruited, 17 SLE patients (14 females, mean age 33.3 ± 7.9 years) and 44 HCs (23 females, mean age 28.3 ± 8.2 years) with high‐quality diffusion and functional MRI data were included in the final analysis (see Table 1).

TABLE 1.

Demographics, ANAM assessment scores, clinical and immunological characteristics.

	SLE	HC	p
Number	22	60	—
Age (years)	34.99 (10.67)	28.43 (8.56)	0.005*
Gender (male:female)	4:18	31:29	0.007*
Handedness (right:left)	20:2	59:1	0.113
Participants with MRI in good quality
Number	17	44	—
Age (years)	33.31 (7.86)	28.38 (8.20)	0.038*
Gender (male:female)	3:14	21:23	0.031*
Handedness (right:left)	15:2	43:1	0.124
WAIS a	94.19 (28.11)	123.21 (20.04)	< 0.001*
Automated Neuropsychological Assessment Metrics (ANAM) a
			Adjusted p value
Code substitution (delayed)	46.32 (16.07)	60.72 (15.89)	0.033*
Code substitution (immediate)	40.04 (14.93)	55.42 (17.41)	0.033*
Code substitution (learning)	47.47 (11.47)	56.45 (10.65)	0.045*
Continuous performance task	71.92 (17.91)	88.61 (19.98)	0.031*
Matching to sample	31.03 (9.19)	42.39 (12.69)	0.031*
Mathematical processing	23.36 (8.45)	28.06 (7.47)	0.058
Spatial processing simultaneous	25.44 (7.08)	30.52 (8.18)	0.079
Simple reaction time	185.94 (28.74)	202.24 (16.06)	0.045*
Memory search	65.89 (17.25)	73.17 (14.99)	0.195
Clinical and immunological characteristics
Disease duration (months)	84.82 (68.80)	—	—
Serum creatinine (μmol/L)	60.38 (12.33)	—	—
Serum albumin (g/L)	38.69 (6.23)	—	—
SLEDAI	4.0 (3.55)	—	—
C3 (mg/dL)	79.35 (32.25)	—	—
C4 (mg/dL)	17.19 (15.04)	—	—
Anti‐dsDNA (IU/mL)	235.94 (257.79)	—	—
Prednisolone use (Yes:No)	17:0	—	—
Daily prednisolone dose (mg/day)	14.85 (12.39)	—	—
Cumulative prednisolone dose (mg)	22.54 (17.81)	—	—
Use of HCQ (Yes:No)	15:2	—	—
HCQ dose (mg/day)	264.71 (122.17)	—	—

Note: Continuous variables are expressed as mean (standard deviation). Individual age and gender were regressed from the ANAM assessment scores before the statistical testing.

Abbreviations: C3: complement component 3; C4: complement component 4; dsDNA: double‐stranded DNA; HC: healthy control; HCQ: hydroxychloroquine; N: number of participants; SLE: systemic lupus erythematosus; SLEDAI: SLE disease activity index.

^a WAIS and ANAM were uncompleted for 2 SLE patients and 1 HC.

^* p < 0.05.

2.2. Clinical and Neuropsychological Assessments

Disease activity of SLE patients was assessed with the SLE Disease Activity Index‐2K (SLEDAI‐2K) (Bombardier et al. 1992), a standard clinical tool used to assess the overall disease activity of SLE using scores ranging from 0 (inactive disease) to > 12 (high disease activity). It evaluates symptoms including rash, joint pain, kidney involvement, seizures, or laboratory test abnormalities to give an overall score, where higher scores indicate more severe disease activity of SLE. As per standard practice of monitoring lupus patients in NUH, sera were assayed for serum complement 3 (C3) and 4 (C4) and anti‐dsDNA levels by immunoturbidimetry and enzyme‐linked immunoassays (BioRad), respectively. Serum albumin and serum creatinine levels are routinely measured as part of standard biochemical profiles. The use of medications, including that of glucocorticoids, was collected from the computerized medical record. All SLE patients received treatments based on their attending rheumatologists' discretion.

IQ was evaluated using the Wechsler Adult Intelligence Scale Fourth Edition (WAIS‐IV) by two certified clinical psychologists (Benson et al. 2010). For the evaluation of neurocognitive performance, the automated neuropsychological assessment metrics (ANAM), a computerized neuropsychological battery, was adopted (Kane et al. 2007; Roebuck‐spencer et al. 2006; Teo et al. 2020). The test evaluates simple reaction time (probing neuromuscular efficiency) and eight domains of neurocognitive assessment, including three code substitution tests (code substitution, delayed code substitution and immediate code substitution, probing learning and recall), spatial processing (probing visual perception and mental rotation), matching‐to‐sample (probing short‐term memory and attention), continuous performance (probing sustained attention and mathematical processing), and memory search (probing working memory). The performance of individual ANAM domains was expressed as individual‐domain total throughput scores (TPS) (Teo et al. 2020).

2.3. Behavioral Task

The participants performed a 4‐min MCST during fMRI scanning, based on the protocol of our previous study (Au et al. 2012; Monchi et al. 2006). They were briefed on the scanning procedures and experimental conditions outside the scanner. The MCST fMRI scanning was performed for two runs, with two blocks in each run. Each block consisted of a total of 24 “switching” trials followed by 12 “identical” trials. In each MCST trial (Figure 1A), four index cards were displayed in fixed order in a row from left to right in the top half of the computer screen: one red triangle, two green stars, three yellow crosses, and four blue circles. A test card was presented below the index cards. Participants were asked to match the test card to the four index cards. During the 24 “switching” trials, the matching was based on one of the three attributes: number, color, or shape. The matching rule was implicit as there was always one matching dimension only with the index cards that the participants had to figure out on their own on a trial‐by‐trial basis. Twelve switch (different rule as previous trial) and 12 nonswitch (same rule as previous trial) trials were pseudo‐randomly interspersed. During the 12 “identical” trials, the test card was an exact match to one of the index cards (i.e., in all three dimensions). Regardless of trial type, in each trial, the test card stayed on the screen until the participant made a key press on a 4‐button response box or for a maximum of 3500 ms, followed by a jittered inter‐trial interval randomly sampled between 500 and 5000 ms. An additional 16 s were included to demarcate the bout of switching trials and identical trials in each block.

FIGURE 1.

Schematic overview of study design. (A) During fMRI scanning, participants performed the Montreal Card Sorting Test (MCST), a cognitive flexibility task in which they matched a stimulus card to one of four index cards based on an implicit rule. (B and D) Individual structural connectivity (SC) networks were constructed from diffusion tensor imaging (DTI) data using a predefined atlas of 144 regions of interest (ROIs), comprising 114 cortical and 30 subcortical regions. (C–F) ROI‐level blood oxygen level‐dependent (BOLD) signals were extracted from the fMRI data and decomposed using graph signal processing into two components: one aligned with the structural network (low spatial frequency) and one liberal from it (high spatial frequency), based on graph Fourier transform. (G) Group comparisons were conducted to assess differences in alignment and liberality between SLE patients and healthy controls, and their associations with task performance were examined. This figure illustrates the study workflow and analytic approach. It does not present empirical results. DA: dorsal attention; DM: default mode; EC: executive control; SM: somatomotor; Sub: subcortical; SVA: salience/ventral attention; Vis: visual.

To examine task switching performance that best characterizes cognitive flexibility, we removed the 12 consecutive identical trials and switch/non‐switch trials with a response time shorter than 200 ms for all blocks. We calculated the accuracy and the mean response time using the remaining trials with accurate answers for each participant as the indicator of MCST performance.

2.4. Image Acquisition

Structural and functional MR images were collected using a 20‐channel head coil in a 3‐Tesla MR scanner (Siemens, PrismaFIT). Functional MRI was performed using a T2*‐weighted echo planar imaging sequence (repetition time = 2000 ms, echo time = 30 ms, flip angle = 90°, field of view = 192 × 192 mm², voxel size = 3.0 mm isotropic, slice thickness = 3 mm, no gap, 36 axial slices, interleaved collection). Diffusion tensor imaging (DTI) was performed using a single‐shot fast‐spin echo planar imaging sequence (repetition time = 5600 ms, 68 slices, field of view = 220 × 220 mm², voxel size = 2.3 mm isotropic, b value = 1000 s/mm², 61 diffusion directions, 8 b0). High‐resolution structural T1‐weighted magnetization prepared rapid gradient echo images (repetition time = 2300 ms, echo time = 2.28 ms, inversion time = 900 ms, flip angle = 8°, field of view = 256 × 256 mm², voxel size = 1.0 mm isotropic) were collected for registration of the diffusion weighted and functional MRI images.

2.5. Image Preprocessing

Task‐based fMRI, diffusion MRI, and structural MRI images were preprocessed using a standard pipeline based on the FMRIB's Software Library (FSL, www.fmrib.ox.ac.uk/fsl) and the Analysis of Functional NeuroImages software program following our previous work (Liu, Poh, et al. 2018; Qian et al. 2018; Wang et al. 2018). The structural image preprocessing included: (1) image noise reduction; (2) skull stripping; (3) linear and nonlinear registration to the Montreal Neurological Institute (MNI) 152 standard space; and (4) segmentation of the brain into gray matter, WM, and cerebrospinal fluid (CSF) compartments.

Preprocessing steps for the task‐based fMRI data included: (1) discarding the first five volumes and interleaved slice‐timing correction; (2) motion correction performed using MCFLIRT (Motion Correction using FMRIB's Linear Image Registration Tool), which realigns each volume to the first functional image (with skull) using rigid‐body transformation; (3) skull stripping; (4) spatial smoothing using a 6 mm full‐width half‐maximum Gaussian kernel to improve the signal‐to‐noise ratio and to reduce inter‐subject variability; (5) structural MRI co‐registration using Boundary‐based Registration and nonlinear registration (FNIRT) to the MNI 152 stereotactic standard space of 2 mm isotropic resolution; (6) nuisance signals reduction by regressing out signals estimated from CSF, WM, global signal, and six motion parameters; and (7) high‐pass temporal filtering (0.009 Hz) to reduce low frequency artifacts.

The DTI data were preprocessed by the following steps: (1) correction for head motion and eddy current distortions using affine registration of each diffusion‐weighted image to the first b = 0 image, implemented with FSL, (2) brain extraction via Brain Extraction Tool (BET), (3) creation of a brain mask with fractional intensity threshold of 0.25 to calculate the maximum motion for each participant by taking the maximum displacement of volumes from the first b = 0 volume within the mask, (4) rotation of diffusion gradients to improve consistency with the motion parameters, and (5) creation of fractional anisotropy (FA) images by fitting a diffusion tensor model to the diffusion data at each voxel. Participants with a maximum motion greater than 3 mm were excluded for further analysis.

2.6. Functional Time Series Extraction and Structural Connectome Construction

For each individual, we extracted the mean BOLD time series concatenated from the first 24 trials in the two blocks of the two runs using a set of 144 regions of interest (ROIs) defined by a previous data‐driven functional parcellation scheme (Thomas Yeo et al. 2011), which defined spatially coherent functional networks using clustering of resting‐state fMRI data from 1000 healthy individuals. The 144 ROIs were grouped into seven intrinsic connectivity networks, namely the salience/ventral attention network (SVN), the dorsal attention network (DAN), the DMN, the ECN, the somatomotor network (SMN), the visual network, and the limbic network. The remaining ROIs covering 30 subcortical regions (Tzourio‐Mazoyer et al. 2002) were grouped into one subcortical network. These networks serve distinct functional roles, ranging from primary sensory processing to higher‐order cognitive functions. For example, the ECN is primarily involved in attention, working memory, and cognitive control, whereas the DMN is associated with internally directed processes such as self‐referential thought, memory retrieval, and creativity (Brown et al. 2019).

We constructed a structural connectivity (SC) network for each participant based on the same predefined set of ROIs, following our previous work (Wang et al. 2018). After generating the fractional anisotropy (FA) images, we used FSL's Bayesian Estimation of Diffusion Parameters Obtained using Sampling Techniques (BEDPOSTX) to model the probabilistic distribution of fiber orientations at each voxel, accounting for potential crossing fibers (Behrens et al. 2003). This step helps to characterize how water diffuses along WM tracts, providing insights into the likely direction and organization of structural brain connections. We then applied FSL's probabilistic tractography tool to the BEDPOSTX output to estimate the likelihood of WM pathways between pairs of ROIs by sampling multiple streamlines from each voxel within a seed region and calculating the proportion of streamlines that reached a target region. This proportion reflects the connection probability between two nodes and forms the basis of the SC matrix. For each node pair ($i$, $j$), the connection probability $P_{\mathit{ij}}$ was defined as the average of the estimated probabilities from node $i$ to $j$ and from node $j$ to $i$. This yielded a symmetric $n\times n$ SC matrix, where each element represents the probability of a structural connection between two brain regions. To account for differences in overall connection density and magnitude across participants (i.e., network cost), and to facilitate inter‐subject comparability, the raw connection probabilities $P_{\mathit{ij}}$ were first log‐transformed to reduce skewness, and then min–max normalized across all connections. The final normalized weight $w_{\mathit{ij}}$ was calculated as:

$$w_{\mathit{ij}}=\frac{\log \left(P_{\mathit{ij}}\right)-\min {\left\{\log \left(P_{\mathit{ij}}\right)\right\}}_{1\le i\ne j\le n}}{\max {\left\{\log \left(P_{\mathit{ij}}\right)\right\}}_{1\le i\ne j\le n}-\min {\left\{\log \left(P_{\mathit{ij}}\right)\right\}}_{1\le i\ne j\le n}}$$ (1)

where $n$ denotes the number of nodes. This individual SC served as the anatomical network upon which the task‐based fMRI signals from the same participant were projected and quantified for degree of alignment and liberality.

2.7. Derivation of Structural–Functional Alignment and Liberality

In classical signal processing, the Fourier Transform decomposes a time‐domain signal into a series of sinusoidal components at different frequencies. Analogously, the graph Fourier transform (GFT) decomposes signals defined over the nodes of a graph (e.g., BOLD activity across brain regions) into components that vary at different “spatial frequencies” across the graph (e.g., the brain structural connectome). These spatial frequencies are defined based on the eigenvectors of the graph's adjacency matrix, which capture patterns of information flow across the structure of the graph. Low‐frequency components correspond to signals that change smoothly across strongly connected nodes, indicating close alignment with the underlying network. In contrast, high‐frequency components represent signals that vary sharply between connected nodes, suggesting deviation or liberality from the underlying network.

In this study, we used GFT to decompose each subject's BOLD signal during the cognitive flexibility task (MCST) into aligned components, which vary smoothly across the structural network and reflect activity closely constrained by anatomical connectivity, and liberal components, which deviate more substantially from the structural network and may reflect more flexible or task‐specific dynamics (Figure 1). We applied the GFT following the previous work (Medaglia et al. 2018). Specifically, the symmetric anatomical network $A\in R^{n\times n}$ ($n$ denotes the number of nodes) had an eigenvector decomposition:

$$A=\mathit{V\Lambda }{V}^T$$ (2)

where $\Lambda$ was the set of eigenvalues, ordered so that ${\lambda }_0\le {\lambda }_1\le \dots \le {\lambda }_{n-1}$, and $V$ was the set of associated eigenvectors $V={\left\{v_k\right\}}_{k=0}^{n-1}$, ordered from “smooth” (i.e., highly connected regions possess values of same signs) to “varied” (i.e., highly connected regions possess values of different signs). We used the eigenvectors $V={\left\{v_k\right\}}_{k=0}^{n-1}$ to define the GFT of the BOLD signal $x\in R^n$ as

$$\tilde{x}=V^{T}x={\left[{\tilde{x}}_0,{\tilde{x}}_1,\dots ,{\tilde{x}}_{n-1}\right]}^T$$ (3)

Given this, we could express the original signal $x$ as

$$x=V\tilde{x}=\sum _{k=0}^{n-1}{\tilde{x}}_k{v}_k$$ (4)

which was a combination of the weighted eigenvector components $v_k$. The contribution of $v_k$ to observed brain signal $x$ was the GFT coefficients ${\tilde{x}}_k$.

We implemented graph filtering to isolate the liberal and aligned components corresponding to the lowest $K_L$ eigenvectors and the highest $K_A$ eigenvectors respectively. The aligned components represent signals that vary smoothly across the graph $A$ (i.e., deviate weakly from structural network) whereas liberal components denote signals that vary highly across the graph $A$ (i.e., deviate greatly from structural network, reflecting reduced anatomical constraints and task‐related flexible processing) at single moments in time. We performed the decomposition at each moment in time and averaged the $K_L$ eigenvectors and the $K_A$ eigenvectors across time to generate the liberality and alignment for each ROI respectively. Intuitively, alignment and liberality measure different amounts of brain functional signal deviation from the underlying anatomical network. Since the selection of $K_L$ and $K_A$ are arbitrary and difficult, we calculated the liberality and alignment for each $K_L$ and $K_A$ ranged from 10 to 40 with a step of one and averaged them across these $K_L$ s and $K_A$ s, respectively.

Given the final ROI‐wise liberality and alignment maps, we calculated the mean liberality and alignment across the whole brain as the global liberality and alignment measures. Similarly, we calculated the mean liberality and alignment across the ROIs within each of the eight networks to get eight network‐level concentrations of aligned and liberal signals. Individuals with higher liberality values exhibit liberal BOLD signal components that deviate more substantially from the underlying anatomical network, whereas lower liberality values indicate that these liberal components are relatively aligned with anatomy. Similarly, higher alignment values reflect that the aligned BOLD signal components are more closely coupled with the underlying anatomy, while lower alignment values suggest weaker coupling between aligned components and the anatomical network.

2.8. Statistical Analyses

Between‐group differences in demographics, cognitive performances measured by ANAM, IQ measured by WAIS, and in‐scanner MCST performance including accuracy and response time were evaluated by using the two‐sample t‐test or the Chi‐Squared test where appropriate. Age and gender were regressed before examining the group difference of ANAM scores and MCST performance.

To examine the group differences in the global and network‐level alignment and liberality measures between SLE and HC, we performed a nonparametric permutation test (alpha level of 0.05) in which we shuffled the memberships across the two groups for 5000 times for each measure. We computed a null distribution by calculating the difference of group mean value from each null model. The measure was judged to be significantly different between the two groups if the real group difference was greater or less than 95% of the null permutations. Age and gender were regressed out from the measures before the permutation tests. In the group comparison of network‐level, multiple comparison correction was performed by controlling false discovery rate (FDR) at p < 0.05.

We then sought to test if the global and network‐level alignment and liberality measures that demonstrated group differences were associated with MCST performance (i.e., accuracy and mean response time) and outside‐scanner cognitive performance (i.e., ANAM subscales) across all the participants and clinical measures within the SLE group using either Pearson's or Spearman's correlation analysis where appropriate. Age and gender were regressed out from the raw measures before the correlation analysis.

To validate the primary findings, we conducted the same analyses on a separate subset of 16 SLE participants and 22 HCs matched on age, gender, and IQ (as operationalized in this study using the WAIS‐IV scale) (SI Table 1).

3. Results

3.1. Demographic, Neurocognitive and Clinical Characteristics

As shown in Table 1, the SLE group was typical of a cohort of SLE patients with clinically stable and inactive diseases. In our sample, patients had low disease activity (mean SLEDAI = 4), indicating clinically inactive SLE status. The SLE group had more females, a higher mean age, and a lower mean overall IQ score than the HC group. The two groups did not differ in handedness. While there was no group difference in MCST performance, including accuracy and response time between SLE and HC, the SLE patients performed inferiorly in most of the cognitive domains as measured by ANAM compared with HC, particularly in sustained attention, concentration, reaction time, working memory, and processing speed.

In the validation samples matched on demographics and IQ (as operationalized in this study using the WAIS‐IV scale), the SLE group was less accurate in the test of matching‐to‐sample, indicating poorer spatial processing and visuospatial working memory (SI Table 1).

3.2. Global Structural–Functional Integration During MCST was Disrupted in SLE and was Related to Cognitive Flexibility

SLE patients showed higher global functional liberality and lower global functional alignment with the structural networks during MCST compared to HC (p = 0.014 and 0.045, respectively; Figure 2), indicating the BOLD signals deviated more strongly from the underlying WM anatomy in SLE. Notably, increased global liberality and reduced global alignment were correlated with reduced task accuracy across SLE patients and HC (r = −0.30, p = 0.020 and r = 0.36, p = 0.006, respectively; SI Figure 1).

FIGURE 2.

Patients with SLE had higher liberality and lower alignment. (A) The mean liberality map averaged across all participants indicated the liberal signals were concentrated largely in visual and somatomotor systems. (B) The mean alignment map averaged across all participants indicated the aligned signals were concentrated largely in subcortical, default mode, dorsal attention, and executive control systems. Patients with SLE showed higher mean liberality from anatomy “*” represents p < 0.05 (panel C) and lower mean alignment with anatomy (panel D) during MCST than HC. Error bars represent standard errors. HC: Healthy control; MCST: Montreal Card Sorting Test; SLE: systemic lupus erythematosus.

Similar results were observed when performing the analyses in the validation samples matched on demographics and IQ (as operationalized using the WAIS‐IV scale) (SI Figure 3).

3.3. ECN Structural–Functional Integration During MCST was Disrupted in SLE and was Related to Cognitive Flexibility

At the network level, SLE patients showed higher liberality within the ECN compared to HC (p = 0.026, FDR‐corrected; Figure 3A,B), whereas no significant group differences in liberality or alignment were observed in other networks (SI Figure 2).

FIGURE 3.

Patients with SLE showed higher functional liberality from anatomy in ECN related to MCST task performance. (A and B) Patients with SLE showed higher mean liberality from anatomy (“*” represented FDR corrected p < 0.05, panel A) and a trend of lower mean alignment with anatomy (panel B) in ECN during MCST than HC. Error bars represent standard errors. (C and D) SLE patients with increased mean liberality and decreased mean alignment in ECN were correlated with reduced task accuracy (lines and statistics in red). The same relationship was observed across all subjects (HC and SLE, lines and statistics in black). Age and gender were controlled in the association analyses. Abbreviations: SLE: Systemic lupus erythematosus. ECN: Executive control network; HC: healthy control; MCST: Montreal Card Sorting Test.

Within the SLE group, we observed that greater ECN liberality was associated with reduced task accuracy (r = −0.50, p = 0.040; Figure 3C, red line) and prolonged response time (r = 0.52, p = 0.033). Additionally, lower ECN alignment was linked to reduced task accuracy (r = 0.56, p = 0.022; Figure 3D, red line). Given the limited sample size (N = 17) in the SLE subgroup, these correlations should be interpreted as exploratory and hypothesis‐generating. Nonetheless, our results suggested a potentially specific role of the ECN in supporting cognitive flexibility in SLE.

When the analyses were extended to the entire sample (SLE and HC), the same patterns emerged: increased ECN liberality was associated with lower task accuracy across participants (r = −0.33, p = 0.012; Figure 3C, black line), and this association remained statistically significant after controlling for global liberality and alignment (r = −0.26, p = 0.048). Similarly, reduced ECN alignment was associated with lower task accuracy across all participants (r = 0.32, p = 0.013; Figure 3D, black line).

Similar results were observed in the validation samples matched on demographics and IQ, as operationalized using the WAIS‐IV scale (SI Figure 4).

3.4. Fronto‐Parietal ECN Subnetworks Drove the Associations With Cognitive Flexibility

In the previous section, we reported the structural‐functional interaction within the ECN played a vital role when performing MCST. Of note, the ECN consists of three subnetworks in the ROI parcellation scheme used in this study (Figure 4A): subnetworks A and B comprise temporal, prefrontal, and parietal regions, while subnetwork C consists of precuneus and regions along the posterior cingulate sulcus. These subnetworks play different roles interacting with other brain networks to execute a complex cognitive process. ECN‐A and ECN‐B, which comprise the key frontoparietal executive control regions, are more directly involved in externally oriented cognitive control and flexible goal maintenance. In contrast, ECN‐C, which is anatomically closer to the DMN, may play a modulatory or integrative role in balancing internally and externally directed attention. Studies have shown that the relative activity levels of these subnetworks can push the ECN into functional alignment with either the default‐mode (task‐negative) or the dorsal attention (task‐positive) network (Murphy et al. 2020). Therefore, we next examined how the signal concentrations in ECN subnetworks differ between the two groups.

FIGURE 4.

Patients with SLE had abnormal structural‐functional integration in ECN subnetworks related to MCST task performance. Among ECN sub‐networks highlighted in the brain surfaces (panel A), ECN‐A and ECN‐B showed higher mean liberality from anatomy during MCST in SLE than HC “*” represented FDR corrected p < 0.05, panel B), and ECN‐A showed lower mean alignment with anatomy during MCST in SLE than HC (panel C). Increased mean liberality and reduced mean alignment in ECN subnetworks were correlated with reduced task performance across SLE and HC groups (panel D–F). Age and gender were controlled in the association analysis. Error bars represent standard errors. ECN: executive control network; HC: healthy control; MCST: Montreal Card Sorting Test; SLE: systemic lupus erythematosus.

When narrowing down to the three sub‐networks of ECN, we found the abnormality of functional organization atop anatomy in the ECN was driven by the higher liberal signal concentrations in ECN‐A and ECN‐B and lower aligned signal concentration in ECN‐A in patients with SLE (FDR‐corrected p = 0.007, 0.008 and 0.026, respectively; Figure 4B,C). Unlike ECN‐C, which is close to DMN, ECN‐A and ECN‐B are in the typical lateral fronto‐parietal executive control regions. Importantly, these ECN subnetwork‐level liberality and alignment showing group difference were associated with task performance across all participants. Specifically, higher liberality in ECN‐A and ECN‐B and lower alignment in ECN‐A were correlated with reduced task accuracy (r = −0.27, p = 0.040; r = −0.32, p = 0.014; and r = 0.28, p = 0.033; Figure 4D–F). Higher liberality in ECN‐B was also correlated with longer task response time (r = 0.29, p = 0.023).

The predominant roles of ECN‐A and ECN‐B in the structural–functional integration disruption in SLE were also validated in the validation samples matched on demographics and IQ, as operationalized using the WAIS‐IV scale (SI Figure 5).

3.5. Disrupted ECN Structural–Functional Integration was Associated With Outside–Scanner Cognitive Performance

We found that higher liberality in ECN was associated with poorer matching‐to‐sample performance, which reflected poorer spatial processing and visuo‐spatial working memory, across all participants (r = −0.41, FDR‐corrected p = 0.013; Figure 5A). When controlling for global liberality and alignment, ECN liberality still accounted for the variance significantly (r = −0.40, FDR‐corrected p = 0.016).

FIGURE 5.

Structural–functional integration in ECN and its subnetworks was related to cognitive performance across all the participants. Higher liberal signal concentration in ECN, ECN‐A, and ECN‐B and lower aligned signal concentration in ECN‐A were associated with poorer matching‐to‐sample performance across all participants (Panels A–D). All the p‐values were FDR corrected. Age and gender were controlled in the association analysis. SI Table 2 showed other associations of ECN structural‐functional integration with other ANAM task performances. ANAM: automated neuropsychological assessment metrics; ECN: executive control network; HC: healthy control; M2S: matching‐to‐sample; SLE: systemic lupus erythematosus.

Consistently, at the ECN subnetwork level, higher liberality in ECN‐A and ECN‐B, and lower alignment in ECN‐A were associated with poorer performance of matching‐to‐sample across all participants (r = −0.37, −0.38, 0.32, respectively and FDR‐corrected p = 0.021, 0.028, 0.043, respectively; Figure 5). The structural‐functional integration in ECN subnetworks was also associated with other ANAM task performances (see SI Table 2 for details).

These associations were also observed in the validation samples matched on demographics and IQ, as operationalized using the WAIS‐IV scale, especially the association of ECN liberality with matching‐to‐sample performance (SI Table 3).

3.6. Disrupted ECN Structural–Functional Integration Was Associated With SLE‐Related Clinical Features

Furthermore, we observed that the higher liberality and lower alignment in the ECN‐B network were associated with lower serum albumin levels (r = −0.52, p = 0.039 and r = 0.55, p = 0.027, respectively) and higher daily prednisolone doses (r = 0.57, p = 0.016 and r = −0.48, p = 0.049, respectively) in SLE participants (Figure 6).

FIGURE 6.

Structural‐functional integration in ECN and its subnetworks was related to SLE‐related clinical features in patients with SLE. Higher concentration of liberal signals and lower concentration of aligned signals in ECN‐B were associated with lower serum albumin level (Panels A and B) and higher daily prednisolone dose (Panels C and D) in SLE participants. Age and gender were controlled in the association analysis. ECN: executive control network; HC: healthy control; SLE: systemic lupus erythematosus.

4. Discussion

In the current study, we employed the MCST, a task that specifically engages cognitive flexibility, and demonstrated disrupted brain functional–structural integration during task performance, especially within the ECN, in SLE patients without clinically overt neuropsychiatric manifestation. Importantly, disrupted functional‐structural integration was associated with poorer cognitive flexibility task performance, out‐of‐scanner cognitive performance, and SLE‐related clinical measures.

While clinically overt neuropsychological manifestations have been defined in the ACR nomenclature and case definitions and the classification criteria for SLE (Liang et al. 1999; Petri et al. 2012), SLE‐related CNS manifestations, particularly neurocognitive function, are often subtle and detectable only through comprehensive neuropsychological tests (Hanly et al. 2012; Tay and Mak 2017). Intriguingly, brain structural abnormalities such as WM volume reduction (Cannerfelt et al. 2018; Mak et al. 2016) and functional alterations, including dysfunctional neural pathways (Liu, Cheng, et al. 2018; Ren et al. 2012; Yu et al. 2019) have been detected in SLE patients despite their clinically subtle or absent CNS symptoms. The MCST task used during our fMRI scanning specifically engages cognitive flexibility, potentially providing a more precise assessment of the neural mechanisms underlying flexibility and rule adaptation compared to resting‐state fMRI. Although group‐level MCST task performance did not differ significantly between SLE patients and HCs, SLE patients exhibited reduced functional alignment and increased liberality during the task. These disruptions in structural–functional integration were associated with poorer MCST performance, highlighting their relevance to individual differences in cognitive flexibility. Importantly, within the SLE group, greater deviation from WM structure (i.e., higher ECN liberality) was associated with poorer performance on the cognitive flexibility task, reinforcing the relevance of structural–functional disruption to individual differences in task behavior among patients. Consistent with the previous study (Medaglia et al. 2018), our findings indicated that less functional liberality and greater functional alignment with WM network organization were facilitative of better cognitive switching performance and provided beneficial organization that efficiently supports cognitive flexibility. The possible mechanism might be that healthier brains are relatively at a natural advantage to meet switching demands and the functional–structural reliance that benefits switching can be detected when task demands necessitate preparedness for cognitive switches. On the contrary, this switching‐specific relationship between signal and anatomy was weakened in patients with SLE, represented by more deviation of the functional activity from the underlying WM anatomy compared to HC.

When each major brain network was examined in greater detail, we found that during MCST, liberal signals within ECN were significantly more concentrated in patients with SLE compared to HC, which was significantly associated with task performance even after controlling for global liberality and alignment. Convergently, a large body of literature on human functional neuroimaging studies using task‐switching and set‐shifting paradigms advocates a central role for the ECN in supporting executive function and cognitive flexibility (Uddin 2021). However, while the ECN regions are recruited during a task requiring flexible selection of items based on different stimulus dimensions, their interactions with other regions are operative in cognitive flexibility. Particularly, the dynamic functional connectivity between ECN and DMN relates to individual differences in cognitive flexibility (Douw et al. 2016). A recent study revealed that the relative activity magnitudes of two distinct and non‐overlapping subnetworks of the ECN tuned the strength of connection between ECN and DMN to achieve cognitive processes, of which one displayed dynamics that were correlated with the dynamics of the DMN, while the other displayed dynamics that were inversely correlated with those of the DMN (Murphy et al. 2020). This observation prompts further exploration of the roles of ECN subnetworks in the structural‐functional integration abnormality in SLE. Our results indicated that the subnetworks in typical lateral frontoparietal regions rather than the subnetworks in precuneus and posterior cingulate sulcus regions which were close to the DMN drove the reduced reliance of functional organization atop anatomy in the ECN in patients with SLE. More interestingly, the structural‐functional integration in ECN, especially the frontoparietal subregions, was also associated with out‐of‐scanner cognitive performance across the participants, indicating that the ECN‐specific changes in structural‐functional alignment/liberality might play a critical role in other cognitive domains including spatial processing, computational skills, concentration, and working memory.

In our study, we did not observe a correlation between the abnormal brain functional‐structural integration and the indices of disease activity, which was likely because our SLE patients were clinically stable and relatively inactive with a mean SLEDAI around four (Mak et al. 2012). Intriguingly, we found the higher concentration of liberal signal and lower concentration of aligned signal in ECN‐B was significantly correlated with lower mean serum albumin level and higher mean daily prednisolone dose in our SLE patients. While these associations may reflect potential links between ECN‐B network alterations and clinical indicators of disease severity or immunosuppressive treatment intensity, the modest sample size warrants cautious interpretation. These findings should be considered preliminary and require validation in larger cohorts. Serum albumin levels have previously been shown to be inversely associated with SLE disease activity (Yip et al. 2010). Although low albumin is often regarded as a surrogate marker of systemic inflammation, the overall low disease activity in our sample and the lack of association between SLEDAI and brain signal abnormalities suggest that inflammation alone is unlikely to account for the observed increase in ECN‐B liberality. An alternative explanation involves the pharmacokinetics of glucocorticoids: at low to moderate doses, serum albumin binds circulating corticosteroids, limiting their free (unbound) fraction and thereby restricting their ability to cross theblood‐brain barrier and influence CNS function (Chau and Mok 2003). Accordingly, lower serum albumin levels may result in a higher proportion of unbound glucocorticoids, increasing the likelihood of CNS effects. This hypothesis is further supported by our previous work, in which we found that higher daily prednisolone doses were associated with poorer performance on mathematical processing tasks, a domain of the ANAM battery that probes executive function and working memory, in patients with SLE (Teo et al. 2020). Consistent with these prior findings, the current study further demonstrated that higher daily prednisolone dose was significantly associated with increased liberality and reduced alignment in ECN‐B during cognitive flexibility task performance, suggesting a possible mechanistic link between corticosteroid exposure and altered structural‐functional brain organization.

A few limitations exist in this study. Firstly, the HC group had a higher IQ than the SLE group; thus, the IQ scores could potentially impact neurocognitive performance and confound interpretation. Nevertheless, the SLE group showed abnormal brain structural‐functional interactions compared to the IQ‐matched HC in our validation analysis, which were also associated with behavioral performance. Secondly, in this cross‐sectional study, the sample size of the SLE participants was modest due to strict exclusion criteria and the quality control of both fMRI and DTI data. Future studies with larger samples are needed to validate and further investigate the observed brain–behavior and clinical associations. Thirdly, a future longitudinally designed study could be conducted to examine the longitudinal changes in brain structural‐functional integration and the effect on cognitive performance. Fourthly, recent work has demonstrated that brain dynamics support flexibility (Uddin 2021); therefore, a future study on time‐varying or dynamic changes in functional alignment with WM anatomy during cognitive switching could seek to further understand the mechanism of cognitive flexibility. Lastly, our previous study identified increased WM extracellular free water in SLE patients (Qian et al. 2022), prompting further investigation into the mechanisms by which WM microstructural abnormalities may disrupt brain structure–function integration in this population (Hu et al. 2024; Silvagni et al. 2021). Future research should also explore other potential contributing factors, including altered functional dynamics, aberrant microglial activation, and blood–brain barrier dysfunction, all of which may play critical roles in the neuropathophysiology of SLE (Bai et al. 2025; Cabrera et al. 2024; Hanly et al. 2023; Nikolopoulos et al. 2023; Wang et al. 2024).

Taking together, by combining the brain functional BOLD signals and WM connectivity, we extended our understanding of the mechanism of cognitive dysfunction in SLE, suggesting that the way the WM connectome constrains the functional activity during cognitive switching is disrupted mainly in the ECN. Our findings provide further insights into the alterations that take place in the structure–function integration relevant to cognitive flexibility in SLE.

Conflicts of Interest

Associate Editor is co‐author: Juan Helen Zhou is a handling editor of “Human Brain Mapping” and a co‐author of this article. To minimize bias, they were excluded from all editorial decision‐making related to the acceptance of this article for publication.

Supporting information

Data S1. Supporting Information.

Qian, X. , Bassett D. S., Ng K. K., et al. 2025. “Disrupted Structure–Function Integration in Systemic Lupus Erythematosus and Its Impact on Cognitive Flexibility.” Human Brain Mapping 46, no. 14: e70299. 10.1002/hbm.70299.

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.