Bayesian stroke modeling details sex biases in the white matter substrates of aphasia

Abstract

Ischemic cerebrovascular events often lead to aphasia. Previous work provided hints that such strokes may affect women and men in distinct ways. Women tend to suffer strokes with more disabling language impairment, even if the lesion size is comparable to men. In 1401 patients, we isolate data-led representations of anatomical lesion patterns and hand-tailor a Bayesian analytical solution to carefully model the degree of sex divergence in predicting language outcomes ~3 months after stroke. We locate lesion-outcome effects in the left-dominant language network that highlight the ventral pathway as a core lesion focus across different tests of language performance. We provide detailed evidence for sex-specific brain-behavior associations in the domain-general networks associated with cortico-subcortical pathways, with unique contributions of the fornix in women and cingular fiber bundles in men. Our collective findings suggest diverging white matter substrates in how stroke causes language deficits in women and men. Clinically acknowledging such sex disparities has the potential to improve personalized treatment for stroke patients worldwide.

Fig. 1. Spatial topography of ischemic stroke lesions summarized from 1401 patients.

We confirm excellent whole-brain coverage, without appreciable hemispheric differences in the voxel-level stroke topography for the whole cohort (a), or separately either for female patients (b) or for male patients (c). The overall lesion volume did not significantly differ between female and male patients within our cohort (two-sided t test, p = 0.13, cf. Table 1). Additionally, there was no significant difference in the left versus right hemispheric overall lesion distribution in females (two-sided t test, p = 0.23) or in males (two-sided t test, p = 0.63). The topographical distributions of tissue damage caused by ischemic stroke are shown in MNI reference space with z-coordinates indicated above each brain slice. Lesions predominantly overlapped in areas that correspond to the vascular territory of the left and right middle cerebral artery. The majority of lesions were localized in subcortical zones, which entails a disconnection of deep white matter tracts. Radiological view; R/L: right/left.

Fig. 2. Multivariate latent-factor discovery deconvolves unique stroke lesion atoms.

In 1401 neurological patients, we used an unsupervised machine learning algorithm to perform a data-driven discovery of white matter lesion patterns. This sum-of-parts approach (non-negative matrix factorization, NNMF) enabled the identification of archetypical white matter lesion configurations, lesion atoms, extracted from high-resolution brain scans with a total use of lesion voxels across patients. There was no significant difference between the expression of each lesion atom in women and men (cf. Supplementary Table 1). a For each extracted lesion atom (y-axis), the underlying configuration of white matter tracts (x-axis) and their contribution from low (purple) to high (yellow) relevance are visualized from the NNMF factor matrix. Each lesion atom (represented as a row in the matrix) is visualized according to the same color scale in the 3D glass brain renderings (c). The most contributing tracts are labeled (for in-depth visualizations, see Supplementary Fig. 2). The anatomical distribution of the derived lesion atoms corresponded to directly interpretable and biologically plausible lesion patterns that are reminiscent of well-known functional systems. Each atom of coherent white matter lesion was assessed for hemispheric asymmetry (b) based on a lateralization index (LI). This index is computed from the weights of the latent factor loadings by LI = (L − R)/(L + R). Hemispheric asymmetry effects are indicated by left (Red) or right lateralization (Blue). Most factors showed a left-hemispheric lesion configuration in correspondence with a similar right-hemispheric homolog. Post hoc naming of the lesion atoms was based on the underlying tract contribution with the highest weights (yellow color scale). Four factors were predominantly associated with the language-related pathways (c): the dorsal perisylvian circuitry (Factor 3 and 9; R and L, respectively) and the ventral pathway (Factor 2 and 5; R and L, respectively). Notably, the core perisylvian pathway shows significant leftward asymmetry in factor 9 with the accentuated influence of the long and anterior AF. Instead, the right homolog in factor 3 predominantly relies on the anterior segment of the AF. Motor function (c) was most strongly expressed in factors 1 (L) and 7 (R), and 10 (bilateral cerebellar network). Primarily limbic factors 4 and 6 were dominated by the influence of the left and right cingulum. Factor 8 was uniquely left lateralized and captured the main effect of the fornix (c). As such, even subtle differences in lateralization align with the current neuroanatomical knowledge of the language pathways. AF arcuate fasciculus, ILF inferior longitudinal fasciculus, UF uncinate fasciculus, IFOF inferior fronto-occipital fasciculus, CST cortico-spinal tract, IC internal capsule, CPC cortico-pontine-cerebellar tract, ICP inferior cerebellar peduncle, SCP superior cerebellar peduncle, OR optic radiation, R/L right/left.

Fig. 3. Stroke lesion atoms are ranked from strongest to weakest associations with clinical outcomes.

We assessed the overall clinical relevance of each lesion atom using Pearson’s correlation between subject-specific lesion atom expression and a cognitive test across outcome measures. a Based on the summed absolute contributions, factors 5 (Ventral pathway L), 8 (Fornix L), and 9 (Dorsal pathway L) ranked as the three most relevant lesion atoms. Following the top three lesion atoms, the limbic factors 4 and 6 reflect the right and left influence of the cingular fiber bundles. b All top five lesion atoms (y-axis, ranked by absolute contribution) showed pronounced negative associations with cognitive impairment across most neuropsychological assessments. Regarding language, damage in the left-hemispheric lesion atoms (Factors 5 and 9) of the ventral and dorsal pathway causes a significant degree of deterioration in the Boston naming test (BN), semantic and phonemic fluency (SF, PF). However, the association with their respective right-hemispheric homologs diverged: the right-hemispheric homolog of the dorsal core perisylvian network (Factor 3) showed a slight negative correlation in BN, but otherwise positive correlations in SF and PF. In contrast, the right-hemispheric homolog of the ventral pathway (Factor 2) is associated with a considerable deterioration in BN and SF. K-MMSE Korean-Mini Mental State Examination, BN Boston Naming, SF semantic fluency, PF phonemic fluency, SVL Seoul-Verbal Learning:, RCFT Rey Complex Figure Test, TMT A/B Korean-Trail Making Test Version A/B, R right, L left.

Fig. 4. Performance of the inferred Bayesian analytical solution in predicting naming and speech fluency.

We confirmed a well-fitted approximation of the underlying distribution as performed for model checking in previous research^,. The posterior predictive checks are shown for the Bayesian models that were estimated to predict the inter-individual difference in (z-scored) Boston naming (a, BN), semantic, and phonemic speech fluency (b, SF, and PF). These model-based simulations of new data were then compared to the actually observed data to compute the overall explained variance (coefficient of determination, R²). This practical check of model-based outcome predictions is a well-recognized approximation to external validation given the actual data at hand. The full model for the BN outcome explained a total variance of R² = 0.51, and the Bayesian models for the PF and SF explained R² = 0.37, respectively. Considering our white matter results, the predictive performance for naming and speech fluency matches previous research on gray matter lesions due to stroke.

Fig. 5. Sex bias becomes apparent from contributions of specific lesion atoms to explaining language outcomes.

For each lesion atom ranked by their clinical relevance (cf. Fig. 3), we show the predictive role in explaining language function separately for women and men. Figure 5 shows the mean of each marginal Bayesian posterior distribution for each lesion atom. Of note, all values reflect a certain degree of incurred dysfunction in language. However, higher mean values correspond to a lesion atom contributing to relatively more preserved language function (red) relative to the rest of our stroke cohort, while lower means explain a relatively more severe loss in function (blue) among all stroke patients in our cohort. Across sexes, effects can be similar or diverging. a For the Boston naming (BN), we observed driving loss of function common to both sexes in the left-lateralized ventral pathway (Factor 5), its right-hemispheric homolog (Factor 2), and the right cingulum (Factor 4). Diverging lesion effects between females and males arise in several atoms, including the right dorsal pathway, the left fornix, cingular tract, and the left and right motor projections (Factors 3, 8, 6, 7, and 1 respectively). b Similarly for semantic fluency (SF), the main effect common to both sexes leading to worse performance was in the left ventral pathway (Factor 5) as well. Less influential were common effects in the right ventral pathway, the left cingulum and fornix (Factors 2, 6, 8). Analogous diverging lesion-outcome predictions in the right dorsal perisylvian circuit (Factor 3) were less pronounced than in BN, leading to moderate impairment for both sexes. But lesions in the left dorsal pathway led to worse outcomes specifically for females. We further observed differences between women and men in the lesioned right cingulum (Factor 4). Here, males were predicted to have a more severe impairment. c For phonemic fluency (PF), adverse outcome is primarily explained by the contributions of the left dorsal pathway (Factor 9) and left Fornix (Factor 8) in both sexes. We observed marked differences between sexes for the right-hemispheric dorsal pathway, left and right cingulum, with males performing worse than females (Factor 3, 4, and 6). In contrast, women showed significantly worse outcomes in the lesioned left ventral pathway (Factor 5).

Fig. 6. Shared and diverging lesion effects on language performance outcomes in female and male stroke populations.

Our model formed full posterior Bayesian estimates to infer the explanatory relevance of each lesion atom for the prediction of language outcomes in female (green) and male (blue) stroke patients. Key assessments of post-stroke language performance included the Boston Naming test (BN, a), semantic (SF, b), and phonemic fluency (PF, c). Lesion atoms were considered relevant if the highest density interval (HDI) of the posterior distribution covering 80% certainty (black line) did not overlap with zero (vertical dotted line). Deviance from zero indicates the robust directional effect (arrow): either a relative loss in function (−) or a relative preserved function (+), compared to all stroke patients in our sample. We found lesion pattern effects that were common to both sexes (left side) and that diverged between women and men (right side). All shared effects indicate a loss in function (a, b, and c: indicated by (−)) for both females and males, while the diverging lesion effects differ in direction between sexes. Here, all lesion atom effects for men indicate worse language function (indicated by − in a–c). For females, instead, effects in BN and SF show a more preserved function (indicated by + in a, b), and a more severe loss in SF (−). Across both sexes, lesion topographies pointed to an overarching structural organization reflecting the dual-stream concept of language. Here, the ventral pathway was critically involved in naming and semantic fluency deficits, and the dorsal pathway in phonemic fluency deficits, congruently for women and men. More importantly, we provide evidence for sex-nuanced lesion-outcome effects. Women primarily rely on left-dominant circuits, including structural connections of the fornix, and a female-specific vulnerability to ventral lesion damage in phonemic fluency. In contrast, men draw from bilateral ventral and cerebellar connections, and potentially integrate more domain-general cognitive and executive functions via right-lateralized cingular fiber bundle to enable efficient language processing after stroke.