A low-dimensional structure of neurological impairment in stroke

Abstract

Neurological deficits following stroke are traditionally described as syndromes related to damage of a specific area or vascular territory. Recent studies indicate that, at the population level, post-stroke neurological impairments cluster in three sets of correlated deficits across different behavioural domains. To examine the reproducibility and specificity of this structure, we prospectively studied first-time stroke patients (n = 237) using a bedside, clinically applicable, neuropsychological assessment and compared the behavioural and anatomical results with those obtained from a different prospective cohort studied with an extensive neuropsychological battery. The behavioural assessment at 1-week post-stroke included the Oxford Cognitive Screen and the National Institutes of Health Stroke Scale. A principal component analysis was used to reduce variables and describe behavioural variance across patients. Lesions were manually segmented on structural scans. The relationship between anatomy and behaviour was analysed using multivariate regression models. Three principal components explained ≈50% of the behavioural variance across subjects. PC1 loaded on language, calculation, praxis, right side neglect and memory deficits; PC2 loaded on left motor, visual and spatial neglect deficits; PC3 loaded on right motor deficits. These components matched those obtained with a more extensive battery. The underlying lesion anatomy was also similar. Neurological deficits following stroke are correlated in a low-dimensional structure of impairment, related neither to the damage of a specific area or vascular territory. Rather they reflect widespread network impairment caused by focal lesions. These factors showed consistency across different populations, neurobehavioural batteries and, most importantly, can be described using a combination of clinically applicable batteries (National Institutes of Health Stroke Scale and Oxford Cognitive Screen). They represent robust behavioural biomarkers for future stroke population studies.

Keywords: behavioural; biomarkers; dimensionality; stroke.

Graphical Abstract

Figure 1

Lesion topography. Overlay of damage in atlas space (n = 164). The colour bar represents the percentage of lesions affecting each voxel (anatomical view).

Figure 2

Behaviour factor analysis. (A) The percentage of variance explained by each Principal Component is proportional to the diameter of each circle. The position of each circle on the brain atlas represents the Principal Component’s lateralization. Finally, circles are labelled with the main subtests underlying each PC, the font size reflects the relative role of each loading. (B) Each table graphically shows single loadings of OCS and NIHSS subtests (on the right) for each PC. Black bars represent positive correlation, while grey bars represent negative correlation.

Figure 3

Correlation matrix of behavioural subtests. The colour bar represents Pearson r-values. Each square corresponds to the variables identified through the PCA analysis (i.e. PC1, PC2, PC3).

Figure 4

Ridge regression scatter plots. Factor scores for right (in green) and left (in red) lesions (in black for midline lesions). The diameter of each coloured circle is proportional to the lesions volume. Each lesion is associated to three Principal Component values: on the X-axis empirically measured behavioural scores, on the Y-axis the estimated behavioural scores of our model. If the lesion location is a good predictor then the relationship between empirical and model scores are linearly related. (A) are calculated on the University of Padua data, while (B) are calculated on the Washington University data.

Figure 5

Ridge regression maps from the University of Padova sample. Warm colours represent positive correlation between anatomical voxels and high PC values (i.e. high level of impairment in the corresponding domains). Cold colours represent negative correlation between anatomical voxels and high PC values. Anatomical overlay maps are shown for PC1, PC2 and PC3 scores, respectively [after Gaussian smoothing (variance = 1) and scaling within $\left[-1,\mathrm{}+1\right]$; weights lower than 0.05 in absolute values are not shown].

Figure 6

Ridge regression maps from the Washington University sample. Warm colours represent positive correlation between anatomical voxels and high PC values (i.e. high level of impairment in the corresponding domains). Cold colours represent negative correlation between anatomical voxels and high PC values. Anatomical overlay maps are shown for PC1, PC2 and PC3 scores, respectively [after Gaussian smoothing (variance = 1) and scaling within $\left[-1,\mathrm{}+1\right]$; weights lower than 0.05 in absolute values are not shown].