Multivariate and network lesion mapping reveals distinct architectures of domain-specific post-stroke cognitive impairments

Background: The purpose of this study was to identify patterns of structural disconnection and multivariate lesion-behaviour relationships associated with post-stroke deficits across six commonly impacted cognitive domains: executive function, language, memory, numerical processing, praxis, and visuospatial attention. Methods: Stroke survivors (n = 593) completed a brief domain-specific cognitive assessment (the Oxford Cognitive Screen (OCS)) during acute hospitalisation. Network-level and multivariate (sparce canonical correlation) lesion mapping analyses were conducted to identify focal neural correlates and distributed patterns of structural disconnection associated with impairment on each of the 16 OCS measures. Results: Network-level and multivariate lesion mapping analyses identified significant correlates for 12/16 and 10/16 OCS measures, respectively which were largely consistent with correlates reported in past work. Language impairments were reliably localised to network-and voxel-level correlates centred in left fronto-temporal regions. Memory impairments were associated with disconnection in a large network of left hemisphere regions. Number processing deficits were associated with damage to voxels centred in the left insular/opercular cortex, as well as disconnection within the surrounding white matter tracts. Within the domain of attention, different subtypes of visuospatial neglect were linked to distinct but partially overlapping patterns of disconnection and voxel-level damage. Praxis impairment was not linked to any voxel-level regions but was significantly associated with disconnection within the left hemisphere dorsal attention network. Conclusion: These results highlight the utility of routine, domain-specific cognitive assessment and imaging data for theoretically-driven lesion mapping analyses, while providing novel insight into the complex anatomical correlates of common and debilitating post-stroke cognitive impairments.


Introduction
Post-stroke cognitive impairments are often conceptualised as being caused by damage to focal brain regions (Moore and Demeyere, 2022;Sagnier et al., 2019;Weaver et al., 2021).However, many common post-stroke cognitive deficits can result from disrupted communication between distant areas (Boes et al., 2015;Gleichgerrcht et al., 2017;Herbet et al., 2015;Mah et al., 2014).These more complex brain-behaviour relationships are important to characterise as they can explain why similar deficits result from a range of different lesion profiles, and can help clinicians develop a more comprehensive understanding of how stroke lesions impact cognition (Bowren et al., 2022;Thiebaut de Schotten et al., 2008).While previous work has reported in-depth analyses of network-level and multivariate anatomy of single neuropsychological deficits, there is a lack of evidence linking performance on brief, multi-domain cognitive screens with their associated complex anatomical correlates.The purpose of the present study was to identify patterns of structural disconnection and multivariate lesion-behaviour relationships associated with impairments across several cognitive domains as captured by the Oxford Cognitive Screen (OCS) (Demeyere et al., 2016).The overarching goal was to determine whether brief, standardised post-stroke cognitive screening measures are able to detect complex brain-behaviour relationships in a diverse and representative sample of acute stroke survivors.
In studies employing traditional voxel-wise univariate lesion mapping, domain-specific post-stroke cognitive impairments have been linked to dissociable patterns of lesion damage (Moore and Demeyere, 2022).This principle has been demonstrated in studies of the neural correlates of single neuropsychological impairments (e.g., Chechlacz et al., 2010;Mirman et al., 2015;Mock et al., 2022;Varjačić et al., 2018), in analyses employing lengthy neuropsychological assessment batteries (Weaver et al., 2021), and by studies using cognitive screening data collected as a component of routine clinical care (Moore and Demeyere, 2022).The use of routinely collected, brief screening data helps avoid many of the key practical limitations associated with lesion mapping studies which rely on detailed neuropsychological batteries (see Moore et al., 2023a;Moore and Demeyere, 2022), at the cost of a degree of behavioural detail.However, past research has demonstrated that lesion mapping analyses which employ behavioural data from brief cognitive screens generally produce results which agree well with lesion mapping studies employing extensive neuropsychological batteries (Moore and Demeyere, 2022).For example, Moore and Demeyere (2022) used univariate voxel-wise lesion mapping to demonstrate that the OCScan reliably differentiate neural correlates supporting 11 distinct cognitive functions using behavioural data collected in a short, standard assessment (<20 min) and routine clinical imaging data.This past work opens up univariate lesion mapping techniques to a wealth of studies that can capitalise on existing clinical data.However, while univariate lesion mapping approaches can effectively identify regions which may contribute to behaviours of interest, they do not always adequately capture the true underlying complexity of brain-behaviour relationships (Herbet et al., 2015;Mah et al., 2014).
Univariate voxel-wise lesion mapping, by design, searches for spatially localised voxel clusters at which the occurrence of lesion damage is associated with a specific behavioural impairment (Bates et al., 2003).This approach can generate accurate results for deficits which are indeed linked to spatially focal neural correlates, but the anatomy of many common post-stroke cognitive deficits is more complex.For example, previous studies have suggested that visuospatial neglect is a disconnection syndrome resulting from non-overlapping lesions impacting diffuse neural networks responsible for distributing attention across space (Bartolomeo et al., 2007;Moore et al., 2023b,d;Saxena et al., 2022).Similarly, praxis impairments have been linked to severed communication between left temporal, parietal, and frontal regions (Hoeren et al., 2014;Rosenzopf et al., 2022).Post-stroke executive function, numerical cognition, language, and memory impairments have all been linked to damage over multiple, spatially distinct regions (Lugtmeijer et al., 2021;Mirman et al., 2015;Moore and Demeyere, 2022;Tsuchida and Fellows, 2013).In cases (such as these) where deficits can result from damage at spatially separated regions, univariate lesion mapping yields results which represent a "spatial average" of all involved correlates, but these may not overlap with the true, underling neural correlates (Mah et al., 2022).
Multivariate and network-level lesion mapping methods were developed to address this critical limitation.Multivariate lesion mapping considers the large-scale spatial distributions of lesion damage across all brain regions simultaneously to identify lesion patterns that distinguish between patients with and without the deficit of interest (Mah et al., 2014;Pustina et al., 2018).Past work has suggested that this multivariate approach yields more precise results than traditional univariate approaches, and is able to more accurately characterise the anatomy of deficits caused by damage to spatially distinct brain regions (Pustina et al., 2018).Despite these advantages, past work has demonstrated that multivariate lesion mapping methods can still yield results that are spatially biased (Ivanova et al., 2021;Sperber et al., 2019) and, in some cases, may use features other than the deficit of interest to distinguish between spared an impaired patients (Moore et al., 2023a).
Conversely, network-level lesion mapping identifies specific patterns of structural disconnection for which the extent of damage predicts the severity of behavioural impairment (Boes et al., 2015;Gleichgerrcht et al., 2017;Salvalaggio et al., 2020).This approach is advantageous because similar patterns of network-level disconnection can result from lesions which do not overlap and could therefore not be detected by voxel-wise approaches (Herbet et al., 2015).However, the majority of network-level lesion mapping approaches rely on standard disconnection atlases (rather than in-vivo tractography), which may not capture potentially critical patterns of anatomical variability across individuals (Griffis et al., 2021).Both multivariate and network-level lesion mapping have been applied to elucidate complex brain-behaviour relationships in a number of individual post-stroke cognitive impairments (e.g., Ghaleh et al., 2020;Saxena et al., 2022;Wiesen et al., 2019;Yourganov et al., 2016).Recent work has also leveraged these approaches to investigate the anatomy of distinct cognitive impairments captured by performance on detailed neuropsychological assessments (Bowren et al., 2022).However, these techniques have not yet been applied to characterise the anatomy of cognitive deficits captured by brief, domain-specific cognitive screening tests in acute stroke survivors.
The current study aimed to address this knowledge gap by applying advanced lesion mapping analyses to identify the neural correlates of common, domain-specific post-stroke cognitive impairments in a large and representative sample of stroke survivors.Our goal was to expand understanding of brain-behaviour relationships as measured by standard cognitive screening tools, used in routine clinical practice.Brainbehaviour relationships are most thoroughly characterised by comparing anatomical results across different levels of analysis, meaning that combining multiple, advanced lesion mapping approaches has the potential to provide novel insight into the anatomy of common and debilitating post-stroke cognitive impairments (Herbet et al., 2015;Moore et al., 2023a).Large-scale studies that characterise the correlates of a range of distinct cognitive impairments within single samples are important for expanding theoretical knowledge of brain-behaviour relationships, as well as for establishing the degree of information complexity that can be extracted from routine clinical databoth in terms of brain imaging and clinically used cognitive screening.

Patients
This retrospective analysis included data from 593 stroke patients recruited in the Oxford Cognitive Screening Programme (Demeyere et al., 2016;Milosevich et al., 2024) (Table 1).The Screening Programme studies included a consecutive sample of acute stroke survivors, excluding only patients who were unable to maintain concentration for 15 min or were unable to provide informed (witnessed) consent.The study was conducted in accordance with the Declaration of Helsinki, and the protocol was approved by the National Research Ethics Committee

Table 1
Demographics and stroke characteristics for the study sample.Where relevant, sample means are reported followed by standard deviations in parentheses.Bilateral strokes were classed as single stroke events which crossed the anatomical midline (e.g., pontine strokes).

Behavioural data
Cognitive status was assessed using the Oxford Cognitive Screen (OCS; Demeyere et al., 2015).The OCS is a short, stroke-specific cognitive screen designed for use in acute stroke settings.It yields 16 distinct measures summarising cognitive impairments across commonly impacted cognitive domains: executive function, language, memory, number processing, praxis, and visuospatial attention (Demeyere et al., 2015(Demeyere et al., , 2016)).This measure is designed to be inclusive for patients with common stroke-related impairments including aphasia, visuospatial neglect, primary visual impairments (e.g.hemianopia), and primary motor impairments (e.g.hemiplegia) (Demeyere et al., 2015(Demeyere et al., , 2016)).
The OCS executive function domain is assessed using a modified (shape-based) trail making task.The OCS language domain consists of a picture naming task (images of four objects/animals), a semantics task (follow verbally presented instructions), and a sentence reading task.The OCS memory domain includes a delayed recall of the previously read sentence (verbal recall task), an incidental multiple-choice recall of stimuli and tasks completed earlier in the OCS (episodic recognition task), and 4 items assessing the participant's orientation in time and location (orientation task).The OCS numerical processing domain includes a number tasks in which participants are asked to write three verbally presented numbers (in digits) and a calculation task where participants are asked to solve four simple addition and subtraction equations.The OCS Praxis domain consists of a gesture and handposition imitation task.Finally, the OCS visuospatial attention domain consists of the hearts cancellation task which yields a total score (number of targets identified) as well as egocentric neglect and allocentric neglect scores.Detailed descriptions of OCS subtests and administration procedures are reported elsewhere (Demeyere et al., 2015(Demeyere et al., , 2016)).
Previous research has demonstrated that impairments are correlated across some OCS domains (Milosevich et al., 2024).For example, found that domain-level impairments in numeric processing and memory were highly correlated (R tet = 0.69), whereas those between attention and praxis were not (R tet = 0.08).Such patterns are expected because specific profiles of post-stroke cognitive impairment may be due to non-random variations in stroke location (Corbetta et al., 2018) and pre-morbid cognitive status (e.g.MCI, cognitive decline) (Deary et al., 2009).Despite domain-level correlations, several large-scale validation studies have provided strong evidence that individual OCS subtests are effectively able to tap their respective functions of interest with minimal interference from co-morbid impairments (Bormann et al., 2024;Demeyere et al., 2015;Kong et al., 2016).The OCS has been validated in large and representative stroke samples (Bisogno et al., 2023;Bormann et al., 2024;Demeyere et al., 2015;Huygelier et al., 2020;Robotham et al., 2020) and provides sufficient cognitive detail to link common post-stroke cognitive impairments to distinct patterns of lesion damage (Moore and Demeyere, 2022).
Where relevant, all binary impairment classifications were made based on standard impairment thresholds.To ensure consistency across scores, patient scores were standardised as proportion correct and ranged between 0 (worst possible score) and 1 (best possible score) for each measure.In cases where independent cognitive impairments are quantified using the same metric (e.g., right and left neglect), all scores not representing significant impairment in the deficit of interest were constrained to 1.This scoring method follows the previously reported approach for lesion mapping studies using OCS data (Moore et al., 2023b).

Neuroimaging
Neuroimaging data were collected as a component of routine clinical care (508 CT,85 MRI (74 T2,3 T1,8 FLAIR)).This combination of different imaging types optimises statistical power by increasing sample size and maximising the degree of lesion overlap.Previous studies have found that these imaging modalities yield comparable lesion mapping results (de Haan and Karnath, 2018;Moore et al., 2023c).The utilised lesion dataset has been demonstrated to be of sufficient quality to localise established correlates of post-stroke cognitive impairments in univariate lesion mapping analyses (Moore and Demeyere, 2022).All lesion masks were manually delineated by trained experts in line with the standard protocol reported by Moore (2022).Native-space lesion masks were smoothed at 5 mm full-width at half-maximum in the z-direction, binarized (0.5 threshold), reoriented, warped into 1 × 1 × 1 mm stereotaxic space using Statistical Parametric Mapping (Ashburner et al., 2016) and Clinical Toolbox (Rorden et al., 2012) functions, and were visually inspected for quality.All lesions which were unable to be accurately normalised were excluded from this study.

Network-level lesion mapping
Network-level lesion mapping was conducted to identify patterns of structural disconnection significantly associated with impairment on each OCS measure.To estimate disconnectivity, patient lesion masks were used to estimate parcel-wise disconnectivity across all cortical areas defined in the Schaefer-Yeo Atlas (100 parcels, 7 networks) (Yeh et al., 2018) and 35 subcortical/cerebellar areas defined in the AAL (Tzourio-Mazoyer et al., 2002) and Harvard-Oxford subcortical atlases.These atlases define 7 functional brain networks: Control, Default, Dorsal Attention, Limbic, Somatic Motor, Ventral Attention, and Visual as well as networks connecting subcortical/cerebellar structures (Schaefer et al., 2018).These network parcellations were established through gradient-weighted Markov Random Field modelling combining local gradient and global similarity measures applied to a range of task-related and resting state fMRI datasets (see Schaefer et al., 2018).
The Schaefer-Yeo cortical parcellation was compared with the HCP-842 atlas to quantify disconnection statistics.The HCP-842 atlas is a population-averaged disconnectome atlas derived from the diffusion MRI data of 842 subjects (Yeh et al., 2018).These individual data were used to map individual white matter tractographies and group these into 550,000 trajectories of representative white matter fascicles (streamlines) (Yeh et al., 2018).In cases in which lesions intersect with each of these defined streamlines, the relevant streamline is considered to be disconnected (Griffis et al., 2021).Network-level disconnectivity statistics were generated by calculating the proportion of streamlines that terminate (end or begin) in each pair of parcels that were disconnected (Griffis et al., 2021).This process produced disconnection matrices in which each cell's value represents the proportion of disconnected streamlines (network edges) connecting each pair of defined grey matter parcels (nodes) per patient (135 nodes, 18,225 edges).This disconnection quantification method is standard for use in network-level lesion mapping analyses (Griffis et al., 2021).
For each network edge impacted in at least five participants, regression analyses were conducted to determine whether percent disconnection at this edge was significantly associated with poor performance on each OCS measure.Each regression included only participants who completed the relevant subtest, and included lesion volume as a covariate of no interest.All disconnection statistics were generated using Lesion Quantification Toolkit (Griffis et al., 2021).
Bonferroni corrections for multiple comparisons were applied.In cases where no results survived Bonferroni corrections, less conservative 5% FDR corrections were applied.This flexible approach was employed to balance the advantages of both correction methods in lesion mapping.For example, Bonferroni corrections are extremely conservative but can reliably identify very strong effects in very large patient samples (Moore et al., 2021;Moore and Demeyere, 2022).However, these corrections have a high false-negative risk when effect sizes (or included samples) are smaller (Mirman et al., 2018).To ensure null results were not due to the use of highly conservative corrections, less conservative FDR corrections were applied in cases where no results survived Bonferroni correction.FDR corrections employ alpha thresholds that are more strict than those used in past network-level lesion mapping studies (e.g., p < 0.05, Philippi et al., 2021).FDR corrections were only employed in cases where Bonferroni corrections did not yield significant results because, in cases where effects are large, FDR corrections often indicate that nearly all conducted tests are significant.This flexible correction approach has been employed in previous network-level lesion mapping studies (Moore et al., 2024(Moore et al., ,2023b)).All cases in which FDR corrections were used are reported.All reported R 2 values are adjusted.

Multivariate lesion mapping
Multivariate voxel-wise lesion mapping was conducted using the sparse canonical correlation (SCCAN) approach implemented in LESY-MAP (Pustina et al., 2018).This technique is an optimisation method which identifies multivariate associations between voxel-level patterns in lesion damage and behavioural scores (Pustina et al., 2018).The approach builds a best-fit model that leverages multivariate lesion patterns to predict behavioural scores.The proportion of voxels retained in models (i.e., model sparseness) is determined through a standardised sparseness optimisation procedure.This procedure involves a four-fold model cross validation operation (using a 75% training/25% testing data split) to evaluate the fit of models with varying degrees of sparseness.SCCAN models are only classed as valid if a model's predicted behavioural scores are above chance (p < 0.05) within the cross-validation procedure.This method has been validated for use in lesion mapping and has been shown to reduce spatial misallocation of voxel-level results relative to traditional univariate lesion mapping approaches (Pustina et al., 2018).
Only voxels impacted in at least 5 patients were considered in SCCAN models.Lesion volume was corrected for using the direct total lesion volume control method to weight input voxel values by lesion size in order to provide a balanced control for the effect of stroke severity on impairment probability (Zhang et al., 2014).SCCAN models were constructed and evaluated for each of the 16 considered OCS measures.Notably, SCCAN models produce voxel values between − 1 and +1, with negative values indicating areas that predict the absence of impairments and positive values indicating areas that predict the occurrence of deficits of interest (Pustina et al., 2018).Given that the aim of our study was to identify areas that, when damaged, lead to significant impairments, only significant voxels assigned positive weights are presented and interpreted.Full SCCAN models (including both positive and negative values) are provided in supplementary materials.
The anatomy of all significant voxel clusters was interpreted relative to the Harvard-Oxford Cortical Atlas and Johns Hopkins White Matter Atlas (Mori et al., 2005).These atlases were selected to generate results that are directly comparable with previous lesion mapping studies (e.g., Moore and Demeyere, 2022), and to provide anatomical results that are balanced in their complexity and ease of interpretation.

Data availability statement
All anonymised data, materials, and code associated with this project are openly available at https://osf.io/depjt/,with the exception of lesion masks.Lesion masks are available upon request from the authors.

Behavioural results
Table 2 presents OCS test results for the included patient sample.Across all considered measures, 86.3% of patients exhibited at least one cognitive impairment, and 72.8% of patients had impairments on two or more OCS measures.Larger lesions were associated with worse performance across all OCS measures (all p < 0.001), with the exception of the trail-making task (p = 1.87), verbal recall (p = 0.278), and right neglect (egocentric p = 0.646, allocentric p = 0.721) (see Supplementary Table 2).Fig. 2 presents lesion overlays for patients impaired on each measure.

Network-level lesion mapping
Network-level lesion mapping analyses identified significant network edges associated with cognitive impairment on 12 of the 16 considered OCS measures.Fig. 3 presents all network edges that were significantly associated with each OCS measure.Fig. 4 summarises the proportion of significant network edges which belong to each of the seven functional networks defined by the Schaefer-Yeo Atlas.
Within the language domain, Picture Naming impairment was associated with damage to 172 edges spanning left frontal and temporal

Table 2
Results from the Oxford Cognitive Screen (OCS).'Tested' refers to the number of participants assessed on each measure.'Number impaired' is the number and percentage of tested participants meeting impairment criteria.'Spared' and 'Impaired' refer to the mean standardised score and standard deviation (in parentheses) for spared and impaired patients on each measure, respectively.regions (R 2 values = 0.002-0.005,all p < 3.64 × 10 − 5 ).Damage to the edge connecting the left hemisphere dorsal attentional network (posterior division 1) and the left default network (parietal division 2) was the strongest predictor of Picture Naming impairment.Impairment on the Semantics task was associated with disconnection across 79 edges within left hemisphere temporo-parietal regions (R 2 values = 0.001-0.002,all p < 3.63 × 10 − 5 ).Impairment in the Semantics task was most strongly predicted by disconnection between the left hemisphere visual network (node 7) and the left hemisphere default network (parietal division 2).
Sentence Reading impairment was associated with damage to 172 left hemisphere edges (R 2 values = 0.002-0.005,all p < 2.98 × 10 − 5 ), with damage to the edge connecting the left hemisphere frontal eye fields (dorsal attention network) and the left putamen being the strongest predictor of impairment.Across all three Language domain tasks, 18.2% of significant network edges were common across all three tasks, with an additional 44.2% of significant edges being common across at least two tasks.
Within the Memory domain, Episodic Recognition deficits were associated with damage to 103 edges mainly located in left temporo-parietal regions (R 2 values = 0.001-0.003,all p < 3.61 × 10 − 5 ).Impairment on the Episodic Recognition task was most strongly predicted by damage to connections between posterior division 1 of the dorsal attention network and parietal division 2 of the default network.Orientation impairment was significantly associated with disconnection of 140 edges (R 2 values = 0.0005-0.001,all p < 0.018 (FDR corrected)).Disconnection between the left hemisphere prefrontal cortex (division 7, default network) and the left pallidum was the single strongest predictor of Orientation impairment.Verbal Recall deficits were associated with damage to 173 edges (R 2 values = 0.002-0.005,all p < 3.26 × 10 − 5 ).The strongest predictor of impairment on this task was damage to streamlines connecting the left frontal eye fields (dorsal attentional network) and the left putamen.Across all Memory domain subtests, 19.2% of all significant edges were common across all three tasks.An additional 50.7% of significant nodes were shared across at least two tasks.
Number Writing deficits were associated with disconnection in 133 edges (R 2 = 0.002-0.006,all p < 3.28 × 10 − 5 ).Damage to the edge connecting the left thalamus and the left frontal opercular cortex (division 2, ventral attention network) was the strongest predictor of impairment.Calculation impairment was linked to disconnection of 9 edges (R 2 = 0.001-0.002,all p < 2.74 × 10 − 5 ).Damage to streamlines connecting the left pallidum and the left somatic motor network (division 4) was the strongest predictor of poor Calculation performance.All network edges significantly associated with Calculation impairment were also associated with Number Writing impairment.
Low Cancellation Total scores (regardless of the presence of response asymmetry) were associated with disconnection in 14 edges (R 2 = 0.002-0.003,all p < 3.54 × 10 − 5 ).Cancellation Total impairment was most strongly predicted by disconnection between the right hemisphere precuneus (control network) and the precuneus/posterior parietal cortex (default network).
Left Egocentric Neglect was associated with disconnection across a wide network, spanning the majority of the right hemisphere (n = 415, R 2 = 0.001-0.002,all p < 3.65 × 10 − 5 ).Left Allocentric Neglect was associated with damage to 107 network edges, spanning right hemisphere temporo-parietal regions (R 2 = 0.0003-0.0006,all p < 3.47 × 10 − 5 ).Disconnection within streamlines connecting the right somatic motor network (division 6) and the dorsal attention network (posterior division 2) were the strongest predictors of both Left Egocentric Neglect and Allocentric Neglect impairments.There was substantial, but not complete, overlap between the networks associated with left egocentric and allocentric neglect.Specifically, 90.7% of edges associated with left allocentric neglect were also significantly associated with left egocentric neglect.The edges associated with allocentric, but not egocentric neglect primarily involved connections between the control and dorsal attentional network (15%), within the control network (10%), and between the ventral attention and somatic motor networks (10%).
Right Egocentric Neglect was associated with damage to 43 edges (R 2 = 0.0004-0.0004,all p < 0.016 (FDR corrected)), with disconnection between the left hemisphere temporal pole (limbic network) and left prefrontal cortex (default network) being the strongest predictor or impairment.Right Allocentric Neglect was linked to disconnection at 16 edges (R 2 = 0.0001-0.0002,all p < 3.31 × 10 − 5 ).Disconnection between the left somatic motor networks (divisions 1 and 2) was the best predictor of Right Allocentric Neglect impairment.Only 3.4% of network edges associated with Right Egocentric Neglect and Right Allocentric Neglect were significantly associated with both deficits.
Visual Field Task impairment was associated with damage to 111 network edges (R 2 = 0.001-0.003,all p < 3.67 × 10 − 5 ).Damage to streamlines connecting the right visual network (division 6) and the right precuneus (control network) was the strongest predictor of impairment on this task.Praxis impairment was linked to disconnection in 60 edges (R 2 = 0.0007-0.002,all p < 0.017 (FDR corrected)), with damage to connections between the left default network (parietal division 2) and the left pallidum being the strongest predictor of Praxis impairment.Within the Executive Function domain, no network edges were significantly associated with poor Trail Making Task performance (FDR corrected).

Multivariate lesion mapping results
SCCAN lesion mapping yielded significant models for 10 of the 16 considered measures.Fig. 5 visualises all voxel clusters that, when damaged, were associated with impairment.Table 3 reports the anatomy of each of these significant clusters.Detailed SCCAN results, including optimum sparseness values and full voxel weight maps, are available in supplementary materials.
Within the Language domain, Picture Naming impairment was linked to damage to large voxel clusters (n = 79,960), spanning left hemisphere anterior temporal regions.Sentence Reading impairment was associated with a more restricted voxel cluster (n = 5130), primarily impacting the   2.
(external capsule) and right middle/superior temporal gyrus.

Discussion
Taken together, our findings suggest that brief, domain-specific cognitive assessment data are able to capture complex network-level and multivariate brain-behaviour relationships in post-stroke cognitive impairment.The identified correlates generally agree well with those documented in past studies that used lesion mapping methods to identify regions associated with impairments identified by detailed neuropsychological assessments.This finding is important as it demonstrates that even simple behavioural data can be used to elucidate complex brain-behaviour relationships.The results highlight the utility of routinely collected cognitive assessment and imaging data for theoretically-driven lesion mapping analyses, and provide novel insight into the complex anatomical correlates of common and debilitating poststroke cognitive impairments.
OCS Language domain impairments were linked to distinct but overlapping network-and voxel-level neural correlates largely centred in left fronto-temporal regions.These findings are in line with past network-level and multivariate analyses of language-related post-stroke cognitive impairments.Past work has suggested that post-stroke aphasia is a multidimensional deficit which can involve deficits in non-verbal cognition as well as classical language deficits (Schumacher et al., 2019).This conceptualisation helps to account for the large and diverse patterns of structural disconnection that were associated with language impairments in the present study.
Within the memory domain, verbal recall, episodic recognition, and orientation task impairment were associated with large and diverse network-level correlates spanning the majority of the left hemisphere.In the SCCAN analysis, verbal recall and episodic recognition were linked to left fronto-temporal correlates, whereas orientation impairment was not associated with any significant correlates.Past multivariate lesion mapping analyses have linked memory deficits to similar voxel-level correlates (e.g., left fusiform, parahippocampal and insular cortices, and deep frontal white matter; Bowren et al., 2022).Notably, orientation impairment was most strongly predicted by damage to the default mode network edge connecting the left pallidum and prefrontal cortex.Previous functional imaging work has found that signals associated with orientation in time and space strongly overlap with the default network (Peer et al., 2015).This suggestion is in line with our study's finding that a higher proportion of default network edges (relative to any other network) were significantly associated with orientation impairment.
Previous studies have indicated that verbal components of memory tasks rely on left-hemisphere language areas while non-verbal memory functions are dependent on right hemisphere cortical and subcortical areas (Mock et al., 2022).Considered in the context of the present study's result, this suggests that the identified correlates of OCS Memory tasks may largely represent the verbal components of these cognitive functions.These results are unlikely to have arisen because patients with Fig. 3. Visualisation of significant network edges associated with impairment on each OCS measure.Network nodes are represented by white dots, and edge colour represents adjusted R2 value for each comparison.On each slice, node locations are collapsed into two dimensions and plotted in ascending order of edge R2 value so that stronger predictors of impairment are plotted on top.Axial slices are MNI z = 14, coronal are MNI y = − 16, and sagittal are MNI x = 8.Names, MNI coordinates, and lesion mapping statistics for each node are openly available (https://osf.io/depjt/).language impairments were unable to complete the OCS memory tasks.Several large-scale validation studies have demonstrated that even patients with severe language impairments are able to complete OCS verbal recall, episodic recognition, and orientation tasks (Bormann et al., 2024;Demeyere et al., 2015;Huygelier et al., 2020).However, such patients may nevertheless have weaker verbal encoding, leading to poorer memory performance (even though the tasks require only recognition from multiple choice options, rather than free recall).
Notably, many patients with exclusively right hemisphere lesions exhibited impairments on these tasks (see Fig. 1), indicating that the memory domain tasks likely capture both verbal and non-verbal memory functions.This was not, however, reflected in the lesion-mapping results.This is likely due to the comparatively greater lesion overlap of impaired patients with left hemisphere versus right hemisphere lesions.Bias toward clusters impacted by lesions in more patients (ignoring critical correlates impacted in fewer patients) is a wellestablished limitation of univariate lesion mapping analyses (Mah et al., 2014).These results illustrate that such prevalence imbalances may impact the results of multivariate, voxel-wise lesion mapping analyses as well.Lesion mapping analyses simplify complex lesion distributions to identify spatially contiguous regions which are responsible for the largest portion of behavioural variance (Mah et al., 2014;Moore et al., 2023a).This method therefore yields a summary of the most strongly associated regions, rather than capturing the wide variability of underlying lesions which can be associated with any given deficit.This methodological feature is important to consider when evaluating the results of the current study, particularly with regard to cases in which only significant unilateral correlates arose from lesion distributions which were clearly bilateral.
It is also important to note that not all memory impairments identified in stroke survivors are causally related to lesion damage.This is because there is a high comorbidity of degenerative brain conditions within the stroke population (e.g.small vessel disease) (Rost et al., 2022).These neurodegenerative conditions commonly result in memory and executive impairments, meaning that a portion of the memory impairments identified by the OCS may have been present prior to stroke onset and related to overall brain health rather than the stroke-specific lesion.Our analyses cannot effectively distinguish between impairments that were pre-morbid and impairments that followed the stroke.This limitation may account for the documented variability of lesion locations in patients with memory impairments.
Past research has indicated that the anatomy of visuospatial neglect deficits is very complex, with previous studies reporting a wide range of voxel-and network-level correlates as being associated with this Fig. 4. Identity of network edges significantly associated with each OCS measure.Each cell represents connections between a given network (listed on x-axis) and another network (y-axis).In cases where both networks are the same, cells represent connections within a single functional network.Cell colour (and label) denotes the proportion of significant network edges which fall into each category.S/C denotes subcortical and cerebellar network edges.syndrome (Moore et al., 2023b,d;Saxena et al., 2022).The results of the current study are broadly in line with this past work, as left egocentric and allocentric neglect were both linked to damage to large and diverse structural networks, spanning a large portion of the right hemisphere (Moore et al., 2023b;Saxena et al., 2022).In line with past work, left egocentric and allocentric neglect were linked to distinct, but overlapping patterns of network-level disconnection (Moore et al., 2023b).Notably, left egocentric neglect impairment was associated with disconnection of some interhemispheric network edges.This result is in line with previous research reporting that damage to interhemispheric network edges (Moore et al., 2023b,d;Saxena et al., 2022) and key hubs for interhemispheric communication (e.g. the corpus callosum (Bartolomeo et al., 2007;Corbetta and Shulman, 2011b;Heilman and Adams, 2003);) are associated with egocentric neglect.
Right egocentric and allocentric neglect were associated with disconnection within largely non-overlapping left-hemisphere networks.The identified network associated with right egocentric neglect impairment differs from the networks identified in previous studies (e. g., Moore et al., 2023b).This is likely due to differences in how right egocentric neglect severity was quantified.Past studies have used detailed "centre of cancellation" metrics to capture the "centre of mass" of participant responses, whereas the present study used basic OCS asymmetry scores (Rorden and Karnath, 2010).These standard asymmetry scores are not able to capture the level of impairment severity detail represented by centre of cancellation scores but were selected to accomplish this study's goal of linking standard OCS scores with their associated neural correlates.This is the first study to identify network-level correlates of right allocentric neglect.Past lesion studies have indicated that this deficit can arise from a diverse range of non-overlapping lesions (Moore et al., 2021).The results of the current study provide evidence that allocentric neglect, like egocentric neglect, may represent a disconnection syndrome.This possibility, in turn, can help account for why diverse, often non-overlapping lesions have previously been associated with allocentric neglect (Chechlacz et al., 2012).
Cancellation task impairment (regardless of neglect impairment) was linked to a diverse range of neural correlates.In SCCAN analyses, cancellation total was primarily linked to damage within right posterior parietal and occipital areas.This finding likely reflects the fact that one of the most common reasons for patients to fail to identify targets on this task is the occurrence of left neglect or left visual field impairments (Demeyere et al., 2016).These impairments are also linked to comparatively homologous lesion patterns (e.g.lesions to the primary visual cortex), which past work has shown are more likely to be identified as significant in voxel-wise lesion mapping analyses relative to lesion patterns that are more spatially variable (Mah et al., 2014;Moore et al., 2023).These findings were mirrored in network-level analyses in which low cancellation total score was best predicted by damage to visual network edges.In line with past work, visual field impairments were primarily associated with damage to, and disconnection between, primary visual areas (Bridge, 2020;Swienton and Thomas, 2014;Zhang et al., 2006).Notably, multivariate analyses primarily yielded right hemisphere correlates of visual field impairment, despite the fact that a number of dissociable impairments (e.g., inhibition, set-switching, working memory) could result in impairment on the OCS trail making task (Muir et al., 2015;Varjacic et al., 2018;Varjačić et al., 2018).Given that the anatomy of each of these constituent abilities is diverse, it is plausible that performance on this single task may not map onto a distinct pattern of lesion damage (Cipolotti et al., 2016;Moore et al., 2023aMoore et al., ,2024;;Varjacic et al., 2018).In line with this possibility, past research has shown that executive impairment is strongly associated with non-lesion related factors (e.g., general cortical atrophy severity, white matter integrity, pre-morbid cognitive decline) (Hobden et al., 2022(Hobden et al., , 2023;;Kirova et al., 2015;Rost et al., 2022).These results highlight that not all cognitive impairments should be expected to map onto distinct damage profiles in lesion mapping analyses.This is because some deficits may be linked to pre-morbid cognitive decline or non-stroke related brain integrity issues including general cortical atrophy or white matter lesions.Similarly, not all cognitive impairments can be expected to map onto distinct patterns of lesion damage, and not all plausible underlying correlates can be effectively detected by even the most advanced lesion mapping methods (Moore et al., 2023a).These possibilities are important to consider in the context of the null lesion mapping results reported in this, as well as in previous studies.
Considered cumulatively, the results of this study suggest that the OCS is able to tap complex brain-behaviour relationships.This finding further emphasises the value of standard, domain-specific post-stroke cognitive assessment as this practice is able to capture complex functional-anatomical relationships which may be overlooked in popular domain-general cognitive screens (Demeyere et al., 2015(Demeyere et al., , 2016(Demeyere et al., , 2019)).This practice is also valuable for clinically relevant lesion mapping investigations as these studies can employ comparatively simple and widely available cognitive data to complex brain-behaviour relationships.Future studies should aim to investigate the prognostic utility of such complex brain-behaviour relationships in order to determine whether detailed lesion location metrics may help predict patient outcomes, over and above other established clinical and cognitive predictors (Bowren et al., 2022;Milosevich et al., 2024).

Limitations
The results presented in this study highlight some of the limitations inherent in all lesion mapping methodologies, regardless of whether univariate, network-level, or multivariate analyses are used.Univariate voxel-wise analyses are biased toward regions in which the lesions of impaired patients are spatially homologous, regardless of underlying functional structures (Mah et al., 2014).Multivariate lesion mapping approaches may help reduce this bias, but this more advanced approach does not appear to entirely negate this risk (Ivanova et al., 2021;Sperber et al., 2019).Multivariate approaches can, in theory, identify significant regions which are unrelated to the function of interest, but incidentally help distinguish between spared and impaired patients (Moore et al., 2023a).Multivariate analyses are also intended to identify the set of regions which most effectively distinguish between spared and impaired patients, rather than capturing the full lesion location variability present in the lesion distributions of impaired patients (Pustina et al., 2018).Similarly, mass univariate network-level analyses are unable to distinguish between edges which are frequently damaged in deficit-causing lesions and edges which are critically associated with the impairment of interest (Mah et al., 2014).Overall, each lesion mapping method has associated strengths and weaknesses.This investigation aims to combine the relative strengths and weaknesses of multiple methods in order to provide a comprehensive investigation of brain-behaviour relationships.
Ideally, cognitive impairment should be diagnosed based on agreement between multiple independent assessments (Moore et al., 2022;Quinn et al., 2021).As this study involved a retrospective analysis, this was not possible.Importantly, different underlying problems can result in apparent cognitive impairment on any given task.This investigation's large sample size reduces, but does not entirely eliminate, the possibility that this inherent noise may have biased reported anatomical correlates.Past research has suggested that distinct patterns of performance on OCS subtests may be explained by reduced number of underlying dimensions and that these dimensions may map on to distinct patterns of lesion damage (Corbetta et al., 2018;Sperber et al., 2023).This possibility is not explored in the present study but represents an important topic for future research to explore.Despite our study's large sample size, not all potentially relevant brain regions are impacted by a sufficient number of lesions to enable statistical analyses (see Fig. 1).The lesion distribution is representative of the stroke population (e.g.primarily MCA strokes) but has comparatively lower power to detect contributions of neural regions supplied by the anterior cerebral artery, posterior cerebral artery, and anterior cerebral communicating artery.

Fig. 1 .
Fig. 1.Lesion overlay for the Study sample (n = 593).Voxel colour denotes the number of patients with lesions impacting each region.Only regions impacted in at least 5 participants are visualised.MNI z-slices -32 -48 are presented.

Fig. 2 .
Fig. 2. Lesion overlays for patients exhibiting impairment on each OCS measure.Voxel colour represents the number of impaired patients with lesions overlapping at each region.MNI z-slices -32 -48 are presented.Total number impaired on each measure is reported in Table2.

Fig. 5 .
Fig. 5. Significant correlates of poor performance as reported by multivariate lesion mapping.All voxels which, when damaged, were associated with worse performance are presented.Voxel colour represents the model weight assigned by SCCAN analysis, with larger values representing stronger relationships between damaged and impairment.MNI slices -40 -58 are presented.Full SCCAN voxel weight maps are visualised in Supplementary Materials and are openly available at htt ps://osf.io/depjt/.