Bayesian modelling disentangles language versus executive control disruption in stroke

Abstract

Stroke is the leading cause of long-term disability worldwide. Incurred brain damage can disrupt cognition, often with persisting deficits in language and executive capacities. Yet, despite their clinical relevance, the commonalities and differences between language versus executive control impairments remain under-specified. To fill this gap, we tailored a Bayesian hierarchical modelling solution in a largest-of-its-kind cohort (1080 patients with stroke) to deconvolve language and executive control with respect to the stroke topology. Cognitive function was assessed with a rich neuropsychological test battery including global cognitive function (tested with the Mini-Mental State Exam), language (assessed with a picture naming task), executive speech function (tested with verbal fluency tasks), executive control functions (Trail Making Test and Digit Symbol Coding Task), visuospatial functioning (Rey Complex Figure), as well as verbal learning and memory function (Soul Verbal Learning). Bayesian modelling predicted interindividual differences in eight cognitive outcome scores three months after stroke based on specific tissue lesion topologies. A multivariate factor analysis extracted four distinct cognitive factors that distinguish left- and right-hemispheric contributions to ischaemic tissue lesions. These factors were labelled according to the neuropsychological tests that had the strongest factor loadings: One factor delineated language and general cognitive performance and was mainly associated with damage to left-hemispheric brain regions in the frontal and temporal cortex. A factor for executive control summarized mental flexibility, task switching and visual-constructional abilities. This factor was strongly related to right-hemispheric brain damage of posterior regions in the occipital cortex. The interplay of language and executive control was reflected in two distinct factors that were labelled as executive speech functions and verbal memory. Impairments on both factors were mainly linked to left-hemispheric lesions. These findings shed light onto the causal implications of hemispheric specialization for cognition; and make steps towards subgroup-specific treatment protocols after stroke.

Keywords: control; domain-general; lateralization; lesion; speech.

Graphical Abstract

Figure 1

Lesion atoms of stroke patterns reveal unique lesion topologies across the whole brain. Voxel-wise information on stroke-induced lesions from >1000 patients was summarized based on 108 anatomical region definitions from a reference atlas (54 per hemisphere) and derived by dimensionality-reducing pattern learning. Region-wise lesion measures were then compressed into 10 essential lesion-pattern ‘prototypes’ in each hemisphere by capitalizing on non-negative matrix factorization (NMF). A. Our derived 10 brain lesion atoms projected on the cortical surface. The resulting lesion atoms capture biologically plausible lesion pattern topographies. B. Relevance of specific brain regions within each of the 10 lesion atoms (quantified as NMF loading) shows whole-brain coverage with distributed lesions in frontal, temporal and parietal as well as subcortical regions. Lesion atoms 1 and 2 implicate the insular cortex and thalamus, respectively. Lesion atom 3 covers the prefrontal cortex (including the inferior frontal gyrus). Occipital regions are covered by lesion atom 4 (inferior occipital cortex and adjacent inferior temporal and parietal regions) and 5 (superior occipital cortex and heteromodal association regions of the inferior parietal lobe, angular gyrus, supramarginal gyrus). Temporal regions are covered by lesion atoms 6 (superior, middle and inferior temporal cortex) and 9 (fusiform gyrus, hippocampus and parahippocampal regions, as well as precuneus and cuneus in the parietal lobe). Precentral and postcentral regions are included in lesion atoms 7 and 10. Lesion atom 8 includes the basal ganglia. C. Z-scored behavioural performance. Raincloud plots show the performance of each subject for the eight cognitive assessments. Please note that the Figure visualizes the input data. We did not perform any statistical tests on these data.

Figure 2

Bayesian hemisphere-aware analysis can robustly predict individual clinical outcomes. A. Model performance of the inferred MIMO (multiple-input multiple-output) Bayesian analytical solution using brain lesion atoms in predicting cognitive impairments. Posterior predictive checks were performed for the Bayesian models that were estimated to predict interindividual differences in eight cognitive measures (on z-scale). Model-based simulations (y-axis) were compared to the observed data (x-axis) to compute the overall explained variance (coefficient of determination, R²). B. Posterior probabilities for the hemisphere-specific lateralization effects for each clinical endpoint. Lesions to the left hemisphere were more relevant for single-patient predictions of cognitive outcomes in the Mini Mental State Exam, BNT, Seoul Verbal Learning, verbal fluency tests, and DSCT. In contrast, damage in the right hemisphere was more explanatory for performance in the RCF and TMT. The posterior model parameters correspond to the upper hemisphere level of our Bayesian multilevel modeling strategy that uncovered the hemisphere-specific model certainty for each cognitive performance dimension. Surface projections: blue colours: lesions are associated with relatively stronger impairments, red colours: lesions are associated with relatively weaker impairments. Std., standard deviation. C. The covariance matrix quantifies the co-occurrences between cognitive outcome measures across patients. Covariance is high (>0.5) for Mini Mental State Exam, BNT and Seoul Verbal Learning, as well as for verbal fluency measures and Digit Symbol Coding Test. In comparison to empirical covariance, the model covariance matrix quantifies the intrinsic relationship between cognitive outcomes given brain lesion conditions and other covariates. D. Similarity of brain-behaviour representations: Hierarchical clustering of model-specific coefficients obtained per clinical endpoint shows distances (similarities) among cognitive measures based on Ward’s method. A first cluster summarizes both fluency assessments and Digit Symbol Coding Test. Based on distances, RCF and TMT can also be grouped. Finally, Seoul Verbal Learning, Mini Mental State Exam and BNT show low distances, suggesting high similarity. Overall, our hemisphere-aware Bayesian model revealed differences in cognitive outcome predictions that follow the distinction between measures of language and executive function.

Figure 3

Four factors trace out the similarity and dissimilarity between language-related and executive function-related clinical endpoints. Low-dimensional representation of clinical endpoints dissociates between language-related and executive control-related factors. A. Scree plot illustrating the cumulative amount of explained variance (cum. expl. variance) in outcome measures and associated eigenvalues. According to the screen test criterion, a four-factor solution is an effective low-dimensional data representation that explains a substantial portion of the variation. B. Factor loadings for the included outcome measures. C. Factor biplot depicting participant scores and loading vectors for the first three factors. Loading vectors represent how strongly each characteristic influences the resulting factor. The angle between a pair of vectors corresponds to the correlation between the given characteristics. Participant factor scores are displayed as points in the plane formed by three principal components. Dot colours code for factor scores (bright colour: high score). D. Performance of the inferred MISO (multiple-input single-output) Bayesian analytical solutions in predicting cognitive impairments for each factor. Posterior predictive checks are shown for the Bayesian models that were estimated to predict interindividual differences in (z-scored) factors. Model-based simulations (y-axis) were compared to the observed data (x-axis) to compute the overall explained variance (coefficient of determination, R²). Right panels: Relevance of the posterior parameter distributions (std = standard deviation) of the left and right hemispheres obtained through separate Bayesian hierarchical models dedicated to predicting each factor. Lesions to the left hemisphere were more relevant for single-patient predictions of cognitive outcomes for factors 1, 2 and 4. In contrast, lesions in the right hemisphere were more relevant for outcome predictions for factor 3. The posterior model parameters correspond to the upper hemisphere level of our Bayesian multilevel modeling strategy that uncovered the hemisphere-specific model certainty for each factor. The four-factor solution, in combination with four Bayesian models, successfully untangled the unique contributions of cognitive outcomes along the language versus executive control axis.

Figure 4

Four factor-specific models trace out distinct stroke-induced lesion patterns for single regions across different cognitive dimensions. Left side: Lesion-deficit prediction for each of the four factors. Coloured brains reflect associations of brain regions with lost (blue colour) or preserved (red colour) cognitive functions. Note that factors 1, 2 and 4 show stronger left-hemispheric lateralization, while factor 3 is right lateralized. Right side: 54 parcels per hemisphere were included based on the Harvard-Oxford cortical and subcortical atlas. The predictions of four factors with differential contributions to language versus executive functions are subserved by distinct brain patterns. β, beta value.

Figure 5

Distinct associations of lesion atoms and cognitive factors disentangle language and executive deficits. Results for the four-factor solution. A. Lesion-deficit associations for the different factors and lesion atoms. Factors 1, 2 and 4 show the most prominent brain lesion effects for atom 6 in the left hemisphere (including temporal regions), while factor 3 shows strong implications for atom 4 (summarizing occipital regions) in the right hemisphere. B. Pearson’s correlations (r) between model-specific beta values dedicated to each factor. Left-dominant factors show stronger correlations with each other. Asterisk denotes False Discovery Rate (FDR)-corrected P-values obtained from a spin-permutation test across the whole brain. C. Distinct relationships between strongest lesion atoms and cognitive impairments. Left: lesion atom 4 in the left hemisphere, with the overall strongest load of the left lateral occipital cortex, is the strongest contributor to factor 3. Middle: Overlap and distinct contributions of the four factors, where each factor is represented by the highest absolute atom loading. Right: lesion atom 6 is characterized by strong left-hemispheric contributions from factors 1, 2, 4, with the strongest load of the middle temporal gyrus. L, R, left, right; NMF, non-negative matrix factorization. D. Correlations of factors with maps of the language network and multiple-demand network^, show a relatively stronger overlap of executive speech functions and language with the language network. Executive functions and verbal memory show relatively stronger associations with the multiple-demand network. Please note that we illustrate an exploratory correlation analysis but none of the values reached significance after correction for multiple comparisons (using False Discovery Rate (FDR)-corrected thresholds of P < 0.05). E. Hemispheric relevance depends on key covariates. The plot depicts the interrelation between the relevance of lesion load in each hemisphere (y-axis, left hemisphere: light colours, right hemisphere: dark colours) and marginal posterior parameters of two key covariates (x-axis) in predicting cognitive performance. Years of education had positive effects on all four factors. The strongest effect was on factor 2. Conversely, an increase on the Informant Questionnaire on Cognitive Decline in the Elderly (IQCODE) scale, i.e. a higher pre-stroke cognitive decline, predicted a decrease in scores of factors 1, 2 and 4. Lesion atom-outcome associations uncovered factor-specific language and executive control deficits: a prominent contribution of the lesion atom covering left-hemispheric superior and middle temporal regions to language, executive speech functions and verbal memory and strong implications of right occipital brain regions for executive functions. IQ, Informant Questionnaire scale. β, beta value.