Cognitive decline is common after ischaemic stroke 1,2 but has proved difficult to predict because the effects of stroke are not limited to the primary neuroanatomical location of brain damage. Functional networks based on correlated brain activity, and structural networks based on white matter fibre connections have been extensively studied in stroke 5,6,39 and reflect widespread disruption, despite focal and heterogenous damage caused by stroke. In contrast, there has been much less investigation of SCNs in stroke, despite their potential to clarify patterns of distributed atrophy, or reflect recovery-related plasticity 13,40. We used a novel, data driven method, taking advantage of our unique longitudinal data to examine covariance in the rate of change in SCNs from subacute to chronic stroke. Crucially, we sought to determine if these SCNs and their longitudinal change had cognitive consequences by examining the relationship to cognitive performance and cognitive decline across domains. We show cognitive decline after ischaemic stroke is associated with degeneration of canonical SCNs.
Structural covariance of the default mode, dorsal attention, executive control, salience, memory and language-related networks was associated with cognitive performance in the attention, executive function, language, memory and visuospatial domains, showing an association between topographical network organisation in subacute stroke and cognitive performance. SCNs seeded from known network nodes replicated the topographical pattern of known functionally specific networks. Using a global ‘brain score’ and ‘behavioural score’ estimated from the PLS model, we found a significant correlation showing cognitive impairment associated with more damaged SCNs that we replicated in the independent validation dataset. Structural covariance of major canonical networks is associated with cognitive performance in subacute stroke. Specifically, more damaged SCNs were related to deficits in attention, executive function, language, memory and visuospatial function. Attention was most implicated in this analysis with three tests of attention covarying with the SCN latent variable, compared to a single test in other domains. This may reflect the frequency of attentional impairment seen in subacute stroke, or the overlap of attentional functions across other cognitive domains.
As a further test of whether SCN integrity was associated with cognitive impairment after stroke, we examined covariance in the rate of longitudinal change in major brain networks and related this to changes in cognitive performance between subacute and chronic stroke. Covariance networks mirrored the SCNs derived at the subacute timepoint, suggesting grey matter volume changes from the subacute to chronic phase were occurring within established, domain-specific networks. The greatest changes, in terms of extent of distributed spatial covarying patterns, were in the default mode network and dorsal attention network.
Degeneration of the major SCNs was associated with cognitive decline, specifically in the attention, memory and language domain. These results suggested that faster degradation of SCNs of bilateral dorsal attention, default mode, language-related networks as well as left dorsolateral prefrontal cortex, hippocampus and frontal insula were related to greater decline in attention, language and memory performance. These two findings are important, but should not be surprising, given that attentional deficits are a consistent feature of post-stroke cognitive impairment. Indeed, the most frequently impaired domains after ischaemic stroke are attention, memory and language 41. Up to 70% of patients have impaired speed of processing and attention after stroke. 1,42,43. Similarly, memory problems are a frequent complaint after stroke, with estimates around 23–55% of patients are affected at three months post-stroke and 11–31% affected at one year 44,45. Our work suggests at least some of this attention, memory and language impairment may be driven by widespread degeneration of SCNs from the subacute to chronic phase. Developmental SCN changes 30, as well as in normal aging and neurological diseases 12,13, have been well characterised. They have rarely associated with cognitive measures, and not been well investigated after ischaemic stroke.
What might be the mechanism resulting in widespread SCN changes, across all networks, after ischaemic stroke, even after accounting for age-related degeneration? Stroke may initiate or aggravate neurodegenerative processes above that seen in healthy aging 46. One plausible mechanism for widespread structural changes as the result of focal ischaemic stroke is secondary Wallerian degeneration due to disconnection between brain regions as a result of the stroke 47. If a brain region, or multiple brain regions in the case of complex networks, are disconnected after stroke, there may be degeneration as a result of underutilisation of the disconnected region that results in volume loss 47. Alternatively, stroke may initiate an ischaemic cascade that results in neurodegenerative processes leading to widespread brain atrophy 46,48,49. Given the timescale of the atrophic changes (3-months and 1-year post-stroke) and how widespread they are in nature, a more plausible alternative is that stroke occurred on a background of accelerated atrophy as the result of cerebrovascular burden 9,50. Future work should examine the clinical characteristics that predict widespread SCN degeneration associated with cognitive impairment.
The work should be interpreted in light of its limitations. As is often the case, the heterogeneity observed within stroke cohorts precludes detailed examination of different profiles of cognitive impairments as the sample size of each subgroup was too small for adequate statistical power. We controlled for infarct volume in the analysis, but did not take into account location, again the heterogeneity of the stroke types and infarct locations, make sub-group analyses under-powered. We conducted an independent validation analysis and confirmed our main findings, namely one significant latent variable accounting for most of the variance in the seed PLS models. Fairly equal contributions from all SCNs as well as significant correlations between brain and behavioural scores were observed in the subacute model, which is similar to the findings in the discovery analysis. However, there were some differences related to the SCN profiles and neuropsychological tests that correlated with the latent variable in the longitudinal model. This may be the result of the sample size used in the validation sample (one third of the discovery sample). Alternatively, this might reflect a degree of dynamic change in cognition at this timepoint. Cognition is likely to stabilise as the time from the stroke increases. Future work will examine the longitudinal effects at even longer, likely even more stable time points collected in this protocol (up to five years). As a group, the median stroke severity (as measured by mRS and NIHSS) was relatively mild. Although this may limit generalisability of the finding to cohorts with more severe stroke, it also raises the possibility that SCNs and cognition maybe even more disrupted when stroke is not as mild as in this cohort. Finally, we carefully chose the seeds to derive our structural covariance networks based on the existing literature. Emerging large functional network atlases 51–53 could be employed to facilitate seed definitions to produce structural covariance networks 54. Future work should aim to replicate our findings to ensure it is robust to seed location.