Predictors of post-stroke cognitive impairment using acute structural MRI neuroimaging: A systematic review and meta-analysis

Background: Stroke survivors are at an increased risk of developing post-stroke cognitive impairment and post-stroke dementia; those at risk could be identified by brain imaging routinely performed at stroke onset. Aim: This systematic review aimed to identify features which are associated with post-stroke cognitive impairment (including dementia) on magnetic resonance imaging (MRI) performed at stroke diagnosis. Summary of review: We searched the literature from inception to January 2022 and identified 10,284 records. We included studies that performed MRI at the time of stroke (0–30 days after a stroke) and assessed cognitive outcome at least 3 months after stroke. We synthesized findings from 26 papers, comprising 27 stroke-populations (N = 13,114, average age range = 40–80 years, 19–62% female). When data were available, we pooled unadjusted (ORu) and adjusted (ORa) odds ratios. We found associations between cognitive outcomes and presence of cerebral atrophy (three studies, N = 453, ORu = 2.48, 95% CI = 1.15–4.62), presence of microbleeds (two studies, N = 9151, ORa = 1.36, 95% CI = 1.08–1.70), and increasing severity of white matter hyperintensities (three studies, N = 704, ORa = 1.26, 95% CI = 1.06–1.49). Increasing cerebral small vessel disease score was associated with cognitive outcome following unadjusted analysis only (two studies, N = 499, ORu = 1.34, 95%CI = 1.12–1.61; three studies, N = 950, ORa = 1.23, 95% CI = 0.96–1.57). Associations remained after controlling for pre-stroke cognitive impairment. We did not find associations between other stroke features and cognitive outcome, or there were insufficient data. Conclusion: Acute stroke MRI features may enable healthcare professionals to identify patients at risk of post-stroke cognitive problems. However, there is still substantial uncertainty about the prognostic utility of acute MRI for this.


Introduction
Cognitive problems after stroke are of major concern to stroke survivors and their families. 1 Identifying who is at risk at the time of stroke may enable healthcare professionals to arrange appropriate follow-up, inform patients and their carers, and plan for possible future health outcomes. Individuals at risk of post-stroke cognitive problems could also be targeted for clinical trials with cognitive endpoints.
The cognitive consequences of stroke are conventionally described as post-stroke cognitive impairment (PSCIimpaired performance on a structured cognitive assessment) and the subcategory of post-stroke dementia (PSD-a clinical diagnosis of a cognitive change sufficient to interfere with daily life).
International guidelines for PSCI highlight that there are currently no prediction tools suitable for clinical practice. 2 A survey of 60 UK healthcare professionals reported that respondents were aware that imaging features could predict PSCI, but they did not use these features in clinical practice. 3 Acute stroke neuroimaging could help healthcare professionals to identify who is at risk of PSCI.
Acute stroke computed tomography (CT) brain imaging is routinely performed in clinical practice to determine the cause of stroke. CT brain imaging is inexpensive and quick to perform but has lower resolution than magnetic resonance imaging (MRI). Recently, MRI has become more available for stroke diagnosis in clinical practice. MRI also allows the identification of neuroimaging features such as cerebral microbleeds (CMB) that are rarely visible on CT brain scans. MRI may help identify neuroimaging features associated with post-stroke cognitive problems.
Cerebral small vessel disease (cSVD) is commonly associated with stroke and dementia. 4 Neuroimaging features include white matter hyperintensities (WMH), CMB, lacunes, perivascular spaces (PVS), recent small subcortical infarcts, and cerebral atrophy. 5 Three systematic reviews have described the associations between neuroimaging features and PSD/PSCI. [6][7][8] One review found that stroke survivors with moderate to severe WMH had a two-to-three-fold increased risk in PSD/PSCI. 7 Another review reported that medial temporal lobe atrophy (MTLA) and global atrophy were associated with increased risk of PSCI, 6 and the third review highlighted an association between MTLA, WMH, and PSCI. 8 These reviews included studies that performed brain imaging up to several months after a stroke, which does not reflect what happens in clinical practice. Only one review performed a sensitivity analysis comparing the association between severity of WMH and PSD when identified on CT versus MRI. 7 The reviews did not report the association between acute stroke lesions and post-stroke cognitive outcome. However, a multicohort study of 2950 stroke survivors reported that infarcts in the left thalamus, left frontotemporal lobes, and right parietal lobe were associated with PSCI. 9 Our previous systematic review focused on the prognostic utility of acute stroke CT finding that presence of atrophy, WMH, and pre-existing stroke lesions were associated with a two-to-three-fold increase in risk of PSD, and WMH was associated with a three-fold increased risk in PSCI. 10 MRI is increasingly being used in clinical practice and is recommended for suspected TIA. 11 A similar review focusing on MRI was needed.

Aims
We determined whether features identifiable on brain MRI in acute stroke can predict PSD/PSCI. We included studies that performed MRI at the time of stroke. We extracted data from the published papers. As this review aimed to be directly applicable to clinical practice, we extracted neuroimaging features (acute stroke lesions and pre-existing stroke features) that could be visually rated on acute MR scans (e.g. presence/absence, severity scales, location).

Protocol and registration
We registered the protocol on PROSPERO (CRD42019 128677). The review is reported according to PRISMA guidelines. 12

Eligibility criteria
Eligibility criteria are outlined in Table 1.

Information sources
We designed a search strategy with an experienced librarian, combining terms relating to stroke, dementia/cognitive impairment, neuroimaging, and study type (Supplement 1). We searched electronic databases: Embase (OVID), MEDLINE (OVID), PsycINFO (EBSCO), and Cochrane Central Register of controlled Trials (CENTRAL) from inception to January 2022. We hand-searched the bibliographies of relevant reviews and included studies. We contacted study authors twice if it was not clear when brain imaging or cognitive follow-up were performed. If the authors did not respond, the study was excluded from the review.

Study selection
We imported studies into Covidence software (Veritas Health Innovation Ltd). 13 Two reviewers independently screened title/abstracts and then full text articles, and conflicts were resolved by consensus or by a third reviewer.

Data collection process
We used a modified version of the CHARMS-PF checklist (CHecklist for critical Appraisal and data extraction for systematic Reviews of prediction Modeling Studies, International Journal of Stroke, 18 (5) tailored to Prognostic Factor studies); 14 12 (~50%) of the included articles were extracted by two reviewers. Disagreements were resolved by consensus or by another reviewer. As disagreements for 12 papers were minor, a single reviewer extracted data from the 14 remaining studies.

Data extraction
We used a data extraction proforma (Supplement 2). If multiple papers included the same cohort, we used the study that presented data most relevant to our primary outcome. We extracted raw data, unadjusted and adjusted associations relating to neuroimaging features. Where various models were presented, we favored the model with the greatest number of variables.

Neuroimaging features
We used the STandards for ReportIng Vascular changes on nEuroimaging (STRIVE) classification system to define neuroimaging features: atrophy, cSVD, WMH, lacunes, CMB, PVS, with additional categories of pre-existing stroke lesions (old infarcts or hemorrhages), acute stroke lesions (ischemic or hemorrhage, presence, number and location), and additional neuroimaging features (cortical superficial siderosis (cSS), hemorrhagic transformation, combinations of features). 5

Cognitive outcome
When studies performed cognitive assessments at multiple time points, we extracted data from the latest assessment after stroke. We produced harvest plots and performed meta-analysis only for studies that assessed global cognitive function/dementia.

Harvest plot
These plots present associations between neuroimaging features and PSD/PSCI, after unadjusted or adjusted analysis, as well as the number of patients in each study and risk of bias.

Meta-analysis
We included studies which reported data to allow calculation of unadjusted (OR u ) or adjusted (OR a ) associations in the meta-analysis. We log-transformed the OR and confidence intervals (CI) so that effect sizes were symmetrical around the null value and performed random-effects meta-analyses using the inverse-variance method. Variability due to betweenstudy heterogeneity was quantified with I 2 . Due to heterogeneity between studies (measurement methods and reporting of data), a limited number of studies were suitable for metaanalysis. Where possible we dichotomized severity of neuroimaging features into presence/absence of these features.
We pooled studies that reported either PSD or PSCI, as there was considerable overlap between the definitions of these groups in different studies. We performed separate metaanalyses for studies that reported unadjusted or adjusted ORs.
We performed sensitivity analyses of studies that excluded patients with pre-stroke cognitive impairment/ dementia (post hoc analysis), excluded hemorrhagic strokes (post hoc analysis), followed-up patients at least 6 months after stroke (planned analysis), and used a neuropsychological battery or diagnostic criteria (post hoc analysis). All analyses were performed using RStudio software (3.6.1).

Quality assessment
We used the Quality in Prognostic factor Studies (QUIPS) tool to assess risk of bias. 14

Results
We identified 10,284 records ( Figure 1) and screened 286 full texts. Forty-six papers were eligible for inclusion (Supplement 3). Multiple papers reported the same stroke population. Findings from 26 papers, comprising 27 stroke-populations (N = 13,114, range of average ages = 40-80 years, 19-62% female) are synthesized in this review.  Kandiah et al. 40 contains two stroke cohorts, we refer to the development cohort as Kandiah et al. 41
There was no association between medial temporal lobe atrophy and PSCI ( Figure 2).
Seven studies (N = 9593) reported data relating to presence, number, and location of acute stroke lesions (Supplement 13). 17,25,27,[32][33][34]36 There was no clear association between acute stroke lesions and PSCI/PSD.  Each unit (box) represents a study. Units that lie above the line of association represent a statistically significant association between neuroimaging feature and cognitive outcome. If the unit lies below the line of association the study did not find a statistically significant association between neuroimaging feature and cognitive outcome. The left hand column (pale blue) represents studies that performed unadjusted analysis. The right hand column (gray) represents studies that performed adjusted analysis. Studies did not always perform both unadjusted and adjusted analyses for the same feature. The height of each unit represents the study sample size (y-axis). The color of each unit represents overall risk of bias for each study (green = low; yellow = moderate; red = high). *Unit height not shown in proportion to study size for this study which is much larger than all others included: study size N = 8700 but represented on the figure as N = 870. Individual study data presented in this plot are reported in Supplement 6-14.

Risk of bias
We rated no studies with high overall risk of bias ( Figure 5). Issues with external validity were common due to studies including only specific stroke types (e.g. lacunar stroke, middle cerebral artery lesion only) and excluding more severe strokes. The majority of studies did not clearly report the reasons for loss to follow-up.

Key findings
This systematic review included 27 cohorts of patients with stroke (N = 13,114). Features of cSVD, visible on acute stroke MRI, were associated with PSCI/PSD. The presence of cerebral atrophy, presence and severity of WMH, presence of CMB, and total cSVD score were associated with increased risk of either PSCI and/or PSD. More severe WMH (adjusted), worse cSVD (unadjusted), presence of cerebral atrophy (unadjusted), and presence of CMB (adjusted) were associated with PSCI/PSD in meta-analyses. We did not find associations between other features and PSCI/PSD or there was insufficient evidence to draw a conclusion. Heterogeneity between studies limited the potential to pool data. International Journal of Stroke, 18 (5) We aimed to explore whether routine MRI collected for clinical purposes at the time of stroke also have a use in predicting long-term cognitive impairment. This is the first systematic review to address the question of whether MRI taken at the time of stroke is useful for identifying patients at risk of post-stroke cognitive problems. Previous systematic reviews included studies that performed brain scans up to several months after stroke. In agreement with these reviews, we also found that WMH were associated with poorer cognitive outcome. 7,8 Crucially, our review looked at pre-existing features and acute stroke lesions visually reported at the time of stroke-finding that pre-existing features were more clearly related than acute lesions to cognitive outcomes-and has clinical implications for early identification of patients at increased risk of PSCI.

Strengths and limitations of this systematic review
In order for our findings to be clinically applicable, we only included neuroimaging features that could be assessed by clinicians, and not those using computerized methods which would require specialist facilities, analysis, and extra time. Although we included brain scans performed within 30 days after a stroke, 78% of the included studies performed scans during acute stroke or within 1 week of the stroke. Studies that assessed PSCI often did not attempt to diagnose dementia, meaning that "PSCI" could include people with mild cognitive impairment or those with dementia. We combined studies that assessed either PSCI or PSD in the same metaanalysis. We did, however, include studies which measured PSCI or PSD separately in our harvest plot, showing association with presence of WMH and CMB and PSD. Dementia was the main cognitive outcome of only four of the included studies; therefore, we can draw limited conclusions from these data. Our review was limited to studies written in English, but we did not restrict the search by language; therefore, we are aware that we were unable to include three studies written in Chinese or Japanese.

Strengths and limitations of included studies
Many of the included studies defined neuroimaging features according to STRIVE criteria which helped when synthesizing findings. 5 However, studies used different measurement methods (presence/severity/location) and analysis techniques (unadjusted/adjusted) to assess the association with cognitive outcome (PSD/PSCI/specific cognitive domains).  We assessed risk of bias using the QUIPS tool. 14 We summed the rating for each risk of bias domain (low = 1, moderate = 2, high = 3) to calculate overall risk of bias (low = 1-7, moderate = 8-13, high = 14-18). When domains were scored as unclear, we carefully considered whether this would increase the overall risk of bias. We used the Risk-of-bias VISualization (robvis) web application to visualize our risk of bias assessments. 43  International Journal of Stroke, 18 (5) Most studies were small in size. Several studies also excluded patients who could not provide informed consent, or who had aphasia/communication difficulties; therefore, findings may not be applicable to patients with more severe strokes.

Research implications
To aid with synthesizing neuroimaging features, studies should provide definitions of the neuroimaging features they are measuring (e.g. STRIVE criteria) and use validated scales. Published guidance on reporting location of acute stroke lesions would be advantageous but do not currently exist. To distinguish which neuroimaging features are associated with PSCI (no dementia) compared to PSD, studies could diagnose according to the Diagnostic and Statistical Manual of Mental Disorders (5th ed.; DSM-5) criteria for major and minor neurocognitive disorder, although reporting full results of cognitive and functional tests is also useful.

Clinical implications
In conjunction with other clinical risk factors such as low education, atrial fibrillation, hypercholesterolemia, and prior stroke (Supplement 16), having a structured way of reporting acute stroke brain scans in clinical practice, that is quick to perform, may help healthcare professionals to identify who is at risk of post-stroke cognitive problems. Should it become possible to identify which stroke survivors are at risk of cognitive problems, future studies need to explore how best to communicate this information to patients and their families.

Conclusions
Routinely performed acute stroke MRI may help healthcare professionals to identify which stroke survivors have an increased risk of post-stroke cognitive problems, but overall effect size is small. Understanding whether patients with acute stroke would want to know this prognostic information, and how best to support them, requires further research.

Authors' note
For the purpose of open access, the author has applied a Creative Commons Attribution (CC BY) license to any Author Accepted Manuscript version arising from this submission.

Declaration of conflicting interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.