ARTICLE https://doi.org/10.1038/s41467-021-23492-3 OPEN Outcome after acute ischemic stroke is linked to sex-specific lesion patterns Anna K. Bonkhoff 1 ✉, Markus D. Schirmer1,2, Martin Bretzner 1,3, Sungmin Hong1, Robert W. Regenhardt 1, Mikael Brudfors4, Kathleen L. Donahue1, Marco J. Nardin1, Adrian V. Dalca5,6, Anne-Katrin Giese7, Mark R. Etherton1, Brandon L. Hancock6, Steven J. T. Mocking6, Elissa C. McIntosh6, John Attia 8,9, Oscar R. Benavente10, Stephen Bevan11, John W. Cole12, Amanda Donatti13, Christoph J. Griessenauer14,15, Laura Heitsch16,17, Lukas Holmegaard18,19, Katarina Jood18,19, Jordi Jimenez-Conde20, Steven J. Kittner12, 1234567890():,; Robin Lemmens21,22, Christopher R. Levi23,24, Caitrin W. McDonough25, James F. Meschia 26, Chia-Ling Phuah17, Arndt Rolfs27, Stefan Ropele 28, Jonathan Rosand1,6,29, Jaume Roquer20, Tatjana Rundek30, Ralph L. Sacco30, Reinhold Schmidt28, Pankaj Sharma31,32, Agnieszka Slowik33, Martin Söderholm34,35, Alessandro Sousa13, Tara M. Stanne36, Daniel Strbian 37, Turgut Tatlisumak38,39, Vincent Thijs40,41, Achala Vagal42, Johan Wasselius43,44, Daniel Woo45, Ramin Zand46, Patrick F. McArdle47, Bradford B. Worrall 48, Christina Jern36,49, Arne G. Lindgren50,51, Jane Maguire52, Danilo Bzdok 53,54, Ona Wu6, MRI-GENIE and GISCOME Investigators and the International Stroke Genetics Consortium* & Natalia S. Rost1 Acute ischemic stroke affects men and women differently. In particular, women are often reported to experience higher acute stroke severity than men. We derived a low-dimensional representation of anatomical stroke lesions and designed a Bayesian hierarchical modeling framework tailored to estimate possible sex differences in lesion patterns linked to acute stroke severity (National Institute of Health Stroke Scale). This framework was developed in 555 patients (38% female). Findings were validated in an independent cohort (n = 503, 41% female). Here, we show brain lesions in regions subserving motor and language functions help explain stroke severity in both men and women, however more widespread lesion patterns are relevant in female patients. Higher stroke severity in women, but not men, is associated with left hemisphere lesions in the vicinity of the posterior circulation. Our results suggest there are sex-specific functional cerebral asymmetries that may be important for future investigations of sex-stratified approaches to management of acute ischemic stroke. A full list of author affiliations appears at the end of the paper. NATURE COMMUNICATIONS | (2021)12:3289 | https://doi.org/10.1038/s41467-021-23492-3 | www.nature.com/naturecommunications 1 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-23492-3 S troke affects >15 million people each year1. It is known to originating from two large, independent hospital-based cohorts result in a substantial overall degree of long-term impair- gathering data of 555 (derivation) and 503 (validation) AIS ment across men and women2,3. However, numerous epi- patients in total. We tailor and deploy sex-aware hierarchical demiological studies indicate clinically relevant, sex-related Bayesian models to simulate predictions of AIS severity and to differences in the characteristics of ischemic cerebrovascular elucidate the sex-specific effects of lesion patterns affecting disease4,5. For instance, due to a longer life expectancy, more similar brain regions in women and men. We seek to map the women than men experience a stroke each year6. Expected lesion constellations underpinning female-specific more severe demographical changes, i.e., an aging population, will widen this strokes, potentially indicating sex-specific maps of functional gap further: in the US, projections suggest that ~200,000 more deficits on the one hand and encouraging more sex-aware acute women will be disabled after stroke than men by 20307. stroke treatment decisions on the other. Such a sex-informed Further sex differences relate to women more often presenting acute stroke care has the potential to alleviate the burden of with non-classic stroke symptoms, such as fatigue or changes in disease on an individual patient level, as well as broader and mental status8,9, and having a higher risk of delays in hospital socioeconomically relevant levels. arrival10,11. Also, women feature a higher risk of cardioembolic stroke due to atrial fibrillation12, which may contribute to the Results often-observed higher acute ischemic stroke (AIS) severity in We here present a generative analysis of acute stroke severity, female patients13. This excess in stroke severity in women persists putting a particular focus on sex-specific lesion pattern effects. even after adjusting for their greater age at onset, comorbidities, We successively combined (1) the automated low-dimensional and prestroke level of independence14,15. Importantly, women embedding of high-dimensional DWI-derived lesion information seem to experience more severe strokes despite comparable lesion via non-negative matrix factorization (NMF)22, and (2) prob- sizes in men and women16. In fact, a similar observation of sex- abilistic modeling, based on the latent NMF embedding, to specific lesion volume effects was noted in the case of aphasia, simulate the prediction of acute stroke severity, as measured by where women had a smaller lesion volume threshold to cause the National Institute of Health Stroke Scale (NIHSS)23. We thus aphasia than men17. first determined pivotal, general lesion pattern effects across all Going beyond lesion volume, lesion-symptom mapping studies patients and successively concentrated on similarities and dif- have enriched our understanding of anatomically unique lesion ferences between men and women (sex assessed by patients’ locations underlying specific symptoms post-stroke18–20. In the medical records). We interpreted explanatory relevances on the case of stroke severity, these analyses have determined widespread level of NMF-derived low-dimensional lesion representations, lesions in white matter, basal ganglia, pre- and postcentral gyri, that we call lesion atoms, as well as the same relevances trans- opercular, insular, and inferior frontal regions to be most relevant formed back to the level of the anatomical gray matter brain for a higher stroke severity, especially if affecting the left regions and white matter tracts. hemisphere21. While these lesion-symptom studies have uncov- ered eloquent lesion locations with high spatial resolution, they Stroke sample characteristics. The derivation cohort consisted of have been systematically blind to any potential sex disparities. If 208 female and 347 male patients (n = 555 in total, mean age considered at all, sex was treated as a nuisance variable and (standard deviation (SD)): 65.0(14.8); 38% women, as indicated regressed out prior to the main analysis21. Thus, none of the by patients’ medical records). The main outcome of interest was recently employed analytical approaches in clinical neuroimaging the acute NIHSS-based stroke severity score within the first 48 h allowed for a dedicated, explicit investigation of sex-specific lesion after admission (median(interquartile range): 3(6), Supplemen- pattern effects in relation to continuous outcome scores. tary Fig. 1). More than one-fourth of stroke patients had a history In this work, we aim to design and conduct a lesion-symptom of hypertension (28.1%), 19.5% had a diagnosis of diabetes mel- analysis capable of capturing male- and female-specific lesion litus, 6.3% atrial fibrillation, and 7.6% coronary artery disease. As patterns, underlying stroke severity in a statistically robust and expected based on prior reports12, women showed a higher fre- spatially precise manner to address previous methodological quency of atrial fibrillation than men (9.1% vs. 4.6%, p = 0.05; c.f., constraints. For this purpose, we leverage neuroimaging data Table 1 for further sex-specific numbers). Acute stroke lesion Table 1 Patient characteristics. All participants Women (n = 208) Men (n = 347) Statistical comparison of male (n = 555) and female patients Age 65.0 (14.8) 67.7 (16.3) 63.3 (13.5) p = 0.001* Sex 62% male, 38% female — — — NIHSS 5.0 (5.9) (median(iqr): 3 5.6 (6.6) (median 4.7 (5.5) (median p = 0.09 (6)) (iqr): 3(6)) (iqr): 3(5)) Normalized DWI-derived stroke 13.7 (29.9) (median(iqr): 13.6 (31.8) (median 13.7 (28.7) (median p = 0.98 lesion volume (ml) 1.7(11.6)) (iqr): 1.5(9.7)) (iqr): 1.7(13.2)) White matter hyperintensity 11.5 (13.5) 12.1 (13.3) 11.1 (13.6) p = 0.42 lesion volume (ml) Hypertension 28.1% 29.3% 27.4% p = 0.63 Diabetes mellitus type 2 19.5% 17.8% 20.5% p = 0.51 Atrial fibrillation 6.3% 9.1% 4.6% p = 0.05* Coronary artery disease 7.6% 6.7% 8.1% p = 0.62 Mean (SD) unless otherwise noted. The groups of male and female patients were compared via two-sample t tests or two-sided Fisher’s exact test as appropriate. Asterisks indicate significant differences between men and women. The disproportionate representation of men and women may reflect an undersampling of female patients as frequently observed in randomized clinical stroke trials87, and may largely stem from the noninclusion of elderly and more severely affected female patients. 2 NATURE COMMUNICATIONS | (2021)12:3289 | https://doi.org/10.1038/s41467-021-23492-3 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-23492-3 ARTICLE a Right-hemispheric lesion atoms b Le-hemispheric lesion atoms R L Lesion atom 1 Lesion atom 6 Lesion atom 8 Lesion atom 4 Lesion atom 2 Lesion atom 5 Lesion atom 7 Lesion atom 9 c 40 paents L R Lesion atom 3 2 paents Lesion atom 10 Fig. 1 Archetypical stroke patterns, lesion atoms, as resulting from non-negative matrix factorization-based dimensionality reduction. A data-driven pattern discovery framework enabled the derivation of coherent patterns of stroke lesion topographies directly from the segmented high-resolution brain scans from 555 stroke patients. This unsupervised, multi-to-multi mapping approach led to unique, predominantly right-hemispheric (a) or left- hemispheric (b) lesion patterns. In case of either one hemisphere, individual lesion atoms represented anatomically coherent cortical and subcortical brain regions and respective white matter tracts, and had varying emphases on more anterior, medial, and posterior regions. While subcortical basal ganglia lesions and cortical lesions in anterior and insular regions were combined in a single lesion atom on the left side of the brain, these patterns were characterized by two separate lesion atoms on the right side of the brain. Lesion atom 8 comprised several brain regions in the left hemisphere; however, also comprised the brainstem and, to a lesser degree, the right thalamus. This pattern likely arose since we did not exclude patients with bilateral stroke (n = 21, 38% female). Some lesion atoms did not comprise any substantial contributions from subcortical brain regions and are shown in transparent. The anatomical plausibility of our derived lesion pattern may particularly stem from the positivity constraint of the non-negative matrix factorization algorithm, an advantageous quality that motivated our choice of dimensionality reduction technique. Conversely, alternative matrix factorization algorithms, such as principal component analysis, would have hampered a straightforward interpretation of lesion pattern effects by encoding individual lesions through more incomprehensible additions and subtractions of low-dimensional lesion pattern. c Similarity of lesion patterns across patients. A voxel-wise lesion overlap is visualized on the left-hand side, while the right-hand side presents region-wise frequencies, i.e., the number of how often a specific region was affected. The maximum lesion overlap was localized subcortically and in the proximity of insular regions in the left and right vascular supply territory of the middle cerebral artery. Significant region-wise differences in lesion loads and frequencies did neither arise between men and women nor between left and right hemispheres (Supplementary Data 1 and 2). Source data are provided as a Source data file. volume did not differ significantly between men and women examined general effects across all patients, on the level of lesion (two-sided t test: p = 0.98). Moreover, neither the frequencies of atoms and on the level of individual anatomical brain regions. how often each cortical and subcortical gray matter region or Successively, we refined analyses and integrated an additional white matter tract was affected, nor the numbers of lesioned hierarchy capturing sex-specific effects. We stratified for male and voxels within each region of interest differed significantly between female sex status and scrutinized sex-specific effects of lesion the sex categories, or between the left and right hemisphere (all atoms and anatomical brain regions. Of note, we corrected all of Fisher’s exact tests or two-sidedt tests p > 0.05, Bonferroni- these analyses for the covariates age, sex, stroke lesion volume, corrected for multiple comparisons, c.f., Fig. 1c, Supplementary white matter hyperintensity lesion volume, and relevant comor- Fig. 1, and Supplementary Data 1 and 2). bidities (atrial fibrillation, hypertension, diabetes mellitus, and coronary artery disease). Notably, we included sex as a variable in Anatomy of the extracted lesion atoms in stroke patients. We the model to differentiate between sex differences in stroke reduced the high-dimensional lesion voxel space by first com- severity that were dependent and independent of lesion patterns. puting the number of lesioned voxels within each of 109 cortical If, for example, stroke severity was generally higher in women, and subcortical gray matter regions, as well as 20 white matter without any link to the actual lesion pattern, conceivably due to a tracts. Subsequently, we employed unsupervised NMF to ten final longer delay between symptom onset and hospital admission and distinct lesion topographies, or lesion atoms. Derived low- decreased likelihood of acute treatment administration, this effect dimensional lesion atoms were found to represent anatomically would be represented by the Bayesian posterior distribution of plausible, hemisphere-specific lesion patterns. The centers of this sex variable, but not in the sex-specific lesion atom Bayesian these lesion patterns varied from anterior to posterior and sub- posterior distribution. In contrast, sex-specific lesion atom cortical to cortical regions, and were broadly similar between distributions would indicate interaction effects on the outcome, hemispheres (Fig. 1a, b). Subcortical and cortical lesion patterns i.e., effects that were specific to an individual’s sex and the precise were represented in separate lesion atoms in the right hemi- lesion atom. Lastly, it is important to note that this adjustment for sphere, while they were captured in a joint lesion atom on the left. global sex differences was independent of the exact knowledge or In correspondence to the primary distribution of individual measurements of their causes, i.e., we did not need to include any lesions, most of these lesion atoms related to infarcts in the left information on the delay in hospital admission explicitly. and right MCA-supply territories and to a lesser extent to infarcts in the posterior circulation. The maximum lesion overlap was Lesion atom and regional relevance for stroke severity. Out of localized in left and right subcortical MCA and insular region the ten derived lesion atoms, five atoms possessed a substantial (Fig. 1c). explanatory relevance for acute stroke severity. This relevance was The low-dimensional representation of lesion topographies inferable from Bayesian posterior distributions of lesion atoms served as input for fully probabilistic, hierarchical linear that did not substantially overlap with zero. In the right hemi- regression models to explain acute stroke severity: first, we sphere, the most relevant lesion atom included subcortical NATURE COMMUNICATIONS | (2021)12:3289 | https://doi.org/10.1038/s41467-021-23492-3 | www.nature.com/naturecommunications 3 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-23492-3 Frontal pole Higher a severity b Opercular cortex Superior temporal gyrus Precentral gyrus Lower severity Postcentral gyrus Lateral occipital cortex Nucleus caudatus Putamen Pallidum Thalamus Brainstem Forceps minor Anterior thalamic radiaon Corcospinal Uncinate tract Fasciculus Subcorcal regions and WM tracts Superior longitudinal Corcal regions fasciculus L R Fig. 2 Local brain regions explaining NIHSS-based stroke severity across 555 patients. a Relevant cortical and subcortical gray matter regions, as well as relevant white matter (WM) tracts. Shows collection of marginal Bayesian posterior distributions from the hierarchical model to explain high vs. low symptom severity (NIHSS). Lesions affecting pre- and postcentral gyri, as well as opercular regions of both hemispheres explained a higher stroke severity. Further brain-behavior effects were left-lateralized: multiple regions, including left middle and inferior frontal gyrus, as well as superior and middle temporal explained a higher stroke severity only when affecting the left hemisphere. While bilateral subcortical regions, in general, had substantial effects on stroke severity, the highest weights were assigned to the putamen and caudate, as well as anterior thalamic radiation, corticospinal tract, and inferior fronto- occipital fasciculus of the left hemispheres. b Brain renderings of region-wise relevances in the explanation of NIHSS-based stroke severity. Source data are provided as a Source data file. regions, i.e., thalamus, nucleus caudatus, putamen, and globus superior and middle temporal gyri, insular cortex, and broader pallidus (lesion atom 4: mean of the posterior distribution = 2.18, opercular regions. highest probability density interval (HPDI) of the posterior dis- In summary, we derived stroke severity-linked lesion patterns tribution covering 90% certainty = 1.43–2.99, Supplementary that highlighted the general importance of subcortical gray matter Fig. 2A). In the left hemisphere, the two most relevant lesion regions and white matter tracts, as well as of bilateral cortical atoms were characterized by both subcortical and cortical regions motor regions, and additional left-lateralized cortical regions, (lesion atom 7: posterior mean = 3.76, HPDI = 2.99–4.49; lesion likely underlying language function. atom 8: posterior mean = 4.8, HPDI = 2.89–6.98, Supplementary Fig. 2B). Affected left and right subcortical regions were similar in explaining acute stroke severity, whereas left cortical affected Differences in lesions patterns between men and women. Next, regions additionally included inferior frontal, insular and superior we concentrated on sex differences in eloquent lesion patterns. temporal gyrus regions, as well as the postcentral gyrus. We refined our Bayesian model and introduced a hierarchical Once projected back to the level of individual gray matter structure that allowed lesion atom effects on stroke severity to regions and white matter tracts, similarities and disparities vary by sex. Previous findings suggest a higher stroke severity in between the left and right hemispheres became apparent as well. women in general, the extent of which can neither be sufficiently Subcortical regions, most notably thalamus, nucleus caudatus, explained by their advanced age nor differences in comorbidities, putamen, globus pallidus, and several white matter tracts the prestroke level of independence or lesion volume14–16. (anterior thalamic radiation, corticospinal tract, inferior fronto- Main patterns of explanatory relevances remained similar to the occipital fasciculus, and superior longitudinal fasciculus) preceding analysis across all patients: for both men and women, the explained higher stroke severity, independent of the lesioned same three lesion atoms, that had already emerged in the joint hemisphere (Fig. 2). Likewise, cortical pre- and postcentral, as analyses, had the highest explanatory relevance. The right- well as supramarginal gyrus and parietal regions explained higher hemispheric lesion atom denoted subcortical regions, among others stroke severity in both the left and right hemispheres. In contrast, representing thalamus, nucleus caudatus, putamen, and globus further cortical effects were more pronounced and more wide- pallidus (lesion atom 4: men: posterior mean = 1.89, HPDI = spread in the left hemisphere. These enhanced left-sided effects 1.00–2.90, women: posterior mean = 2.64, HPDI = 1.44–3.82, mainly related to middle and inferior frontal gyri, as well as Fig. 3). The lesion atoms in the left hemisphere combined the 4 NATURE COMMUNICATIONS | (2021)12:3289 | https://doi.org/10.1038/s41467-021-23492-3 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-23492-3 ARTICLE Lesion atom 6 Lesion atom 1 Lesion atom 6 Lesion atom 1 Lesion atom 7 Lesion atom 2 Lesion atom 7 Lesion atom 2 Lesion atom 8 Lesion atom 3 Lesion atom 8 Lesion atom 3 L R L R Lesion atom 9 Lesion atom 4 Lesion atom 9 Lesion atom 4 Lesion atom 10 Lesion atom 5 Lesion atom 10 Lesion atom 5 Frequency Frequency parameter weight parameter weight Fig. 3 Sex-specific Bayesian posterior distributions of all ten lesion atoms and overall whole-brain region-wise relevance to explain stroke severity in women (left) and men (right). Our Bayesian framework was purpose-designed to enable fully probabilistic estimations of the parameters that quantify the associations of the ten lesion atoms with stroke severity. These Bayesian posterior parameter distributions are shown in outer circles, corresponding lesion atom renderings are presented in the subjacent circle (right-hemispheric lesion atoms: shaded in yellow-olive, left-hemispheric lesion atoms: orange- yellow; distributions that substantially diverged from zero are nontransparent). Lesion atoms 7 and 8 of the left hemisphere and lesion atom 3 of the right hemisphere had the highest weights, implying a high relevance in explaining stroke severity, in both men and women. In view of seven relevant lesion atoms in women (lesion atoms: 1, 4, 5, 7–10), yet only five of such relevant lesion atoms in men (lesion atoms: 2, 4, 6–8), lesion patterns were more widespread in women compared to men. This female-specific more widespread pattern becomes additionally apparent in whole-brain renderings of region- wise relevances explaining stroke severity, as visualized in circle centers (c.f., Fig. 4 for details). Source data are provided as a Source data file. same subcortical structures, as well as the brainstem and additional lesion embeddings, likely expressing sample-specific lesion dis- cortical regions, all of them in proximity to the insular cortex in the tributions, lesion atoms of the derivation and validation dataset left hemisphere (lesion atom 7: men: posterior mean = 3.15, HPDI were overall highly correlated. This high correlation indicated = 2.22–4.02, women: posterior mean = 4.76, HPDI = 3.49–6.07; that our unsupervised approach facilitated the derivation of lesion atom 8: men: posterior mean = 2.37, HPDI = 0.393–4.27, similar lesion topography embeddings in both independent women: posterior mean = 7.04, HPDI = 3.75–10.50, Fig. 3). datasets (Supplementary Table 2). In addition, back-projected Women presented with generally more widespread explanatory region-wise relevances were highly correlated when computed relevances as seven out of ten lesion atoms posterior distributions based on the other cohort’s lesion embedding (derivation cohort: did not overlap with zero (Fig. 3, lesion atoms 1, 4, 5, 7–10). In r = 0.84, p < 0.001; validation cohort: r = 0.78, p < 0.001). Both of men, only five lesion atoms posterior distributions did not these correlation analyses combined, thus highlighted the dis- overlap with zero (Fig. 3, lesion atoms 2, 4, 6–8). tillation of largely similar archetypical lesion pattern in both Once projected back to the level of individual brain regions, these cohorts and the independence of results from the concrete lesion more widespread lesion pattern effects in women were also visible, embedding. in particular regarding cortical gray matter regions (c.f., Figs. 3 and The most relevant regions explaining stroke severity were also 4). Manifest differences between men and women emerged for five located subcortically in the left and right hemisphere, as well as in specific lesion atoms: for women, the right-hemispheric lesion atom bilateral precentral and postcentral gyri and left-hemispheric 1 and the four left-hemispheric lesion atoms 7–10 had substantially insular and opercular regions (Supplementary Fig. 4a). Similar to higher explanatory relevances, i.e., the distribution of the difference our analysis before, women presented more widespread eloquent between posterior distributions of male and female patients did not lesion patterns compared to men (Supplementary Fig. 4b, c). In overlap with zero (Fig. 5). Lesion atom 1 was mainly characterized particular, we found substantial differences between men and by right-sided frontal, insular and opercular, as well as precentral women in a lesion atom that predominantly comprised left- regions. Lesion atoms 7–10 represented cortical and subcortical hemispheric, presumably posterior cerebral artery-supplied regions of the entire left hemisphere. regions (lesion atom 10: difference distribution: mean = −2.68, HPDI = −4.92 to −0.733 (i.e., no overlap with zero)). Sex differences relating to this specific left-sided lesion atom thus Validation analyses. Similar main findings were shown when appeared to be the most pronounced and robust. repeating the analyses in an independent, multisite dataset24. Again, we extracted ten lesion atoms that captured typical stroke patterns in the left and right hemispheres in low dimensions Ancillary analyses. We aimed to gain further insights into the (Supplementary Fig. 3). While there were subtle differences of influences of (i) the exact time of imaging and stroke severity NATURE COMMUNICATIONS | (2021)12:3289 | https://doi.org/10.1038/s41467-021-23492-3 | www.nature.com/naturecommunications 5 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-23492-3 a b Lower Higher Lower Higher severity severity severity severity Fig. 4 Local brain regions explaining stroke severity. a Female-specific results (208 women) and b male-specific results (347 men). In both men and women, subcortical lesions affecting gray matter regions and white matter tracts explained higher stroke severity. Similarly, cortical presumptive bilateral motor and left-lateralized language regions (e.g., especially bilateral precentral and postcentral gyri, left-sided inferior frontal and superior, middle temporal gyri) also explained higher stroke severity. In difference to men, women featured more widespread and also more pronounced lesion pattern, including a greater range of cortical regions contributing to stroke severity, e.g., the left superior, middle and inferior temporal gyrus, left angular gyrus and lateral occipital cortex, lingual gyrus, as well as precuneus and parahippocampal cortex gyrus. These differences in eloquent lesion patterns arose despite comparable total lesion volumes for men and women (two-sided t test: p > 0.05). Source data are provided as a Source data file. acquisition, (ii) cardioembolic vs. non-cardioembolic stroke female-specific lesion atom effects independent of whether subtypes, and (iii) potential hormonal effects. comparing men and women with cardioembolic strokes, or While data were acquired within the first 48 h after admission men and women with non-cardioembolic stroke genesis: in case for the entire derivation cohort, we aimed to reduce potential of cardioembolic stroke, difference distributions of lesion atoms 4, confounds due to varying times of imaging and stroke severity 6, and 8 suggested a higher explanatory relevance exclusively in acquisition, as well as any acute revascularization therapy effects women (Supplementary Fig. 9), while lesion atoms 1, 2, 9, by specifically investigating those patients with MRI imaging and and 10 had a substantially higher relevance in women in case of NIHSS score available upon their admission. These criteria were non-cardioembolic stroke (Supplementary Fig. 10). Since we fulfilled by 152 patients in total (men: 87, mean age(SD): 63.1 extracted female-specific effects in strata of both only cardioem- (12.3), mean NIHSS(SD): 4.6(5.6); women: 65, mean age(SD): bolic and non-cardioembolic stroke patients, these results 69.6(14.5), mean NIHSS(SD): 5.5(6.5); two-sidedt tests: age: p = rendered the interpretation of lesion pattern effect differences 0.003, NIHSS: p = 0.45, Supplementary Fig. 6). The Bayesian due to varying frequencies of cardioembolic stroke unlikely. posterior difference distribution of lesion atom 10, representing Furthermore, all of these female-specific effects emerged for presumed left posterior cerebral artery (PCA)-supplied brain samples that were fairly even in their numbers of men and regions, indicated a substantial female-specific effect in explaining women (c.f., cardioembolic stroke: 108 men and 95 women), or a higher stroke severity (Supplementary Fig. 7). did not comprise any significant differences in age and stroke Women, especially if older, are known to more frequently severity (c.f., non-cardioembolic stroke: two-sidedt tests: age: p = experience atrial fibrillation and cardioembolic strokes12. The 0.45, NIHSS: p = 0.44), which additionally increased the con- motivation of our second ancillary analysis was to investigate fidence that sex differences did not artificially arise from these whether observed sex differences in explanatory lesion patterns differences. merely stemmed from varying frequencies of cardioembolic Finally, we aimed to explore the potential effects of sex strokes. As information on stroke subtype was available for hormones, such as estrogen, which are known to be markedly subgroups of each cohort only, we here pooled data from the affected by menopause25. We stratified the entire group of derivation and validation cohort to rest analyses on a sample as patients according to their sex and an age cutoff of 52 years, the large as possible. Approximately one-fourth of the 880 patients in median age at menopause26. All of the female-specific lesion atom total experienced a cardioembolic stroke (203 patients with effects, as observed in the main analysis, disappeared in the cardioembolic stroke: men: 108, mean age(SD): 66.6(14.3), mean analysis of all men and women below the age of 52 years (men: NIHSS(SD): 6.3(5.4), women: 95, mean age(SD): 74.3(12.9), mean 113, mean age(SD): 43.1(8.5), mean NIHSS(SD): 5.4(6.1), women: NIHSS(SD): 7.5(7.0); two-sidedt tests: age: p < 0.001, NIHSS: p = 87, mean age(SD): 42.1(7.9), mean NIHSS(SD): 4.3(5.2); two- 0.17; 677 patients with non-cardioembolic stroke: men: 419, mean sidedt tests: age: p = 0.41, NIHSS: p = 0.19; Supplementary age(SD): 62.8(13.3), mean NIHSS(SD): 4.7(5.3), women: 258, Fig. 11). What is more, lesion atom 7 was now assigned a higher mean age(SD): 63.7(16.4), mean NIHSS(SD): 5.1(5.7); two-sidedt relevance in male patients (Supplementary Fig. 12). In contrast, tests: age: p = 0.45, NIHSS: p = 0.44, Supplementary Fig. 8). As we observed female-specific higher relevances in three lesion expected, women featured a higher frequency of cardioembolic atoms (lesion atoms 1, 9, and 10), when comparing men and strokes (Fisher’s exact test: p = 0.03). We detected substantial women in the subgroup of patients above the age of 52 years of 6 NATURE COMMUNICATIONS | (2021)12:3289 | https://doi.org/10.1038/s41467-021-23492-3 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-23492-3 ARTICLE a Lesion atom 1 b Lesion atom 7 nucleus caudatus frontal pole Difference distribuon Difference distribuon middle & inferior frontal pole putamen frontal gyrus Insular cortex pre-central thalamus gyrus c Lesion atom 8 d Lesion atom 9 angular & Difference distribuon post-central gyrus supramarginal gyrus Difference distribuon post-central gyrus brainstem superior & middle thalamus lateral occipital & inferior temporal gyrus cortex occipital pole e Lesion atom 10 Difference distribuon precuneus Less than 5% of the difference distribuon Plausibility overlaps with zero, indicang a substanal female-specific effect occipital pole thalamus R fusiform more hippocampus gyrus relevant lateral occipital in women cortex visual cortex Fig. 5 Five lesion atoms showed substantially higher relevance for explaining stroke-induced symptom severity in women than men. a Lesion atom 1. Less than 5% of samples of the difference Bayesian posterior distribution between men and women (i.e., posterior distribution of lesion atom 1 in men— posterior distribution of lesion atom 1 in women) overlapped with zero, suggesting a substantially larger lesion atom 1 effect in women. This specific lesion pattern represented right-hemispheric lesions in frontal, insular, and precentral regions (difference distribution: mean = −1.2, HPDI = −1.96 to −0.31). b–d Lesion atoms 7–9. Almost all lesion atoms of the left hemisphere, but the one representing precentral cortex and middle and inferior frontal gyrus regions, indicated more pronounced effects on stroke severity in women. Lesion atom 7 highlighted left-sided insular cortex and subcortical regions (difference distribution: mean = −1.61, HPDI = −3.12 to −0.02). Lesion atom 8 comprised brainstem lesions, as well bilateral, left-hemispherically more distinct thalamus lesions (difference distribution: mean = −4.67, HPDI = −8.55 to −1.08) and lesion atom 9 widespread left-sided cortical lesions relating to the superior, middle and inferior temporal gyri, postcentral cortex, angular and supramarginal gyrus, and latero-occipital cortex (difference distribution: mean = −3.42, HPDI = −6.15 to −0.63). e Lesion atom 10. Lesions in left-sided brain regions of the presumed posterior circulation were associated to more severe strokes specifically in women (difference distribution: mean = −1.99, HPDI = −3.90 to −0.22). The sex difference for lesion atom 10 can be considered the most pronounced and robust one, given it was reliably observable in all ancillary analyses and was replicated in an independent dataset of stroke patients. All analyses were corrected for overall sex effects, i.e., lesion atom-independent effects (e.g., due to potential female-specific delayed hospital arrival). Of note, this correction was independent of the explicit knowledge or measurement of the causes of these global sex differences. Source data are provided as a Source data file. age (Supplementary Fig. 13, men: 533, mean age(SD): 68.4(9.5), effects in the derivation cohort, featuring significantly older mean NIHSS(SD): 4.9(5.2), women: 325, mean age(SD): 73.0 female patients, than in the validation cohort that was (11.0), mean NIHSS(SD): 6.0(6.3); two-sidedt tests: age: p < 0.001, characterized by a nonsignificant age difference between men NIHSS: p = 0.004). Overall, female-specific effects were thus and women. noticeably more pronounced with advanced age. Given that we Of note, female-specific effects relating to lesion atom 10, ascertained female-specific effects for both cardioembolic, as well representing left-hemispheric presumably PCA-supplied regions, as non-cardioembolic stroke subtypes in the previous ancillary were the only ones that consistently emerged in all of the three analysis, the observations for older female patients are unlikely ancillary analyses. Thus, these results confirmed the notion due to an increased frequency of cardioembolic stroke and linked arising from the main derivation and validation analyses that sex lesion patterns. Further, this observable sex–age interaction might differences concerning this lesion atom were the most also explain why we discovered more extensive female-specific reliable ones. NATURE COMMUNICATIONS | (2021)12:3289 | https://doi.org/10.1038/s41467-021-23492-3 | www.nature.com/naturecommunications 7 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-23492-3 Discussion severity, cannot be broken down and thus does not allow for In this study, we combined a probabilistic lesion-symptom conclusions on specific functions, such as verbal and performance mapping technique with empirical lesion data originating from scale IQ. This level of granularity hampers a direct comparison to two large independent cohorts of 555 (derivation) and 503 earlier studies. Nonetheless, we also find that particularly women (validation) AIS patients to derive and validate sex-specific lesion are vulnerable to left-hemispheric lesions. Indeed, we can relate patterns underlying NIHSS-based acute stroke severity. the most robust excess vulnerability of female vs. male patients to Across all patients, lesion patterns highlighted the relevance of anatomically precise lesion locations in the left-hemispheric PCA bilateral subcortical and white matter regions, as well as bilateral territory, specifically featuring hippocampal, thalamic, and pre- cortical motor regions and left-lateralized cortical insular and cuneal regions. Based on existing knowledge on these regions’ opercular regions, likely representing regions underlying language physiological functions, it may be suggested that lesions in these function. This distribution of main weights was rendered parti- regions more likely underlie (higher) cognitive, than, for example, cularly plausible, given that the NIHSS scale assigns the majority basic motor functions. This predilection toward specific anato- of its points to motor and language functions. In addition, results mical regions within the left hemisphere could also explain dif- from both of our cohorts, and those results originating from ferences seen in earlier lesion studies, as these studies stratified established voxel-wise lesion-symptom mapping (VLSM) patients only based on left- or right-sided lesion locations without analyses21 were very similar. This congruency increased the investigating any greater spatial detail of lesions29,30,32. confidence in our methodological approach, as well as in the Altogether, a sex-specific lateralization effect, comparable to accuracy and reproducibility of results. the considerable extent and spatial distribution of the one that we When comparing men and women, eloquent lesion patterns detected in two independent datasets, has not been described in were generally more widespread in female patients, implying that previous lesion studies, nor in studies focused on sex dis- more regions contributed to stroke severity in women. These sex crepancies in functional cerebral asymmetries in healthy adults28. discrepancies were particularly pronounced for lesions affecting a This discrepancy between previous reports and our findings may number of left-hemispheric regions, i.e., thalamus, hippocampus, originate from (i) the greater anatomical resolution in our study and occipital cortical brain regions, thus regions reminiscent of compared to early sex-aware lesion studies, (ii) not explicitly the vascular territory of the posterior circulation. Lesions in these considering sex in any of the recent lesion-symptom regions were underpinning a substantially higher stroke severity studies18,21,33–35, and (iii) the fact that our lesion-symptom in women only, indicating sex-specific distributions of functional analysis jointly investigated effects of functional asymmetry and deficits. the capacity for acute compensation, e.g., by means of brain Our sex-specific findings are of interest from several points of plasticity. view. First, we could not detect any sex differences in total nor- We furthermore observed signs of an interaction effect of sex malized lesion volume, or in any normalized lesion volumes of with age, when stratifying the entire sample based on the median the atlas-based brain regions and white matter tracts. Rather, age at menopause26. None of the female-specific lesion pattern similarly configured lesion patterns were linked to more severe effects could be detected, when comparing men and women strokes in women compared to men. Stroke severity in men was below the age of 52 years. In fact, we rather saw indications for a predominantly explained by five specific lesion patterns, denoting male-specific effect pertaining to left-hemispheric subcortical and bilateral subcortical and left-sided insular and opercular regions. insular brain regions in this age subgroup. In contrast, three Women were characterized by similar, yet more pronounced lesion atoms, among others relating to the left-hemispheric PCA relevant lesion patterns, as well as several additional relevant territory, were found to be more pronounced in men and women ones. These additional lesion patterns did not possess any rele- above the age of 52 years. This constellation was thus suggestive vance in men. Therefore, the female-specific effects can be con- of a possible influence of sexual hormones, that are dramatically sidered more substantial and spatially widespread. Most changed in women after menopause36. noticeably, lesions in the presumed vascular territory of the left Sex differences in brain organization in general and in PCA, among other regions affecting left hippocampal and tha- stroke incidence and outcome in particular are related to the lamic regions, left fusiform, lingual, and intracalcarine cortex, as influence of sex steroid hormones, with estrogen likely being well as left precuneus and cuneal cortex explained a dis- the most prominent one36,37. These hormones are assumed to act proportionally high stroke severity in women compared to men. via irreversible organizational and reversible activational This difference was consistently observable across datasets and mechanisms. The former effect implies the facilitation of definite did not appear to represent an artifact of a more frequent car- male or female tissue phenotypes, while the latter requires the dioembolic stroke genesis, more severe strokes or a higher age at presence of the hormone for an effect25. Any sex difference that stroke onset in women. changes with menopause is thus rather linked to activational It is important to note that the strongest observed sex differ- hormonal pathways. An example of sex differences in stroke ences were strictly lateralized to the left hemisphere. Previous thought to be due to activational hormonal effects is the reduced research suggests that male or female sex and respective sex stroke risk due to the premenopausal cycle of estrogen. This hormones contribute to induce functional cerebral protective effect is lost after menopause38 and may not be rees- asymmetries27. Men appear to have a stronger hemispheric tablished by hormone replacement therapy independent of the asymmetry; however, while robustly replicated, determined effect time of initiation39,40. Experimental research has furthermore sizes have been small28. Such an enhanced asymmetry in men was shown that female animals experience smaller stroke lesions than also found in some early lesion studies on intelligence29. How- male animals41. This effect could be neutralized by ovariectomy ever, further early lesion studies suggested that lateralization and consequently a decrease in estrogen levels42. It was inter- differences between the sexes might be even more complex, i.e., preted as hormone-linked sex-specific sensitivity to cerebral female brains may be asymmetric to a comparable degree, yet in ischemia25,29,43. In fact, male-specific cell cultures of hippo- different ways30,31. In particular, it was found that left- campal neurons and astrocytes seem to be more vulnerable to hemispheric lesions in women led to both verbal and perfor- ischemia than female-specific cell cultures44, and even ischemia- mance scale IQ deterioration, while only one quality—either independent research indicates an important role of estrogen in verbal or performance—was affected in all other lesion and sex sustained hippocampal structural plasticity and associated cog- constellations30,32. Our outcome measure, NIHSS-based stroke nitive function45,46. 8 NATURE COMMUNICATIONS | (2021)12:3289 | https://doi.org/10.1038/s41467-021-23492-3 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-23492-3 ARTICLE The effect of age that we witnessed here—stronger female- ancillary analyses suggested more pronounced female-specific specific effects in older patients, yet none in younger patients effects in case of more advanced age. —suggests an activational nature of the apparent sex Overall, an effective step toward tailored medical stroke disparities. While our outcome measure of global stroke severity care55,56 may lie in more sex-aware acute treatment decisions. does not allow for fine-grained evaluations of implicated func- Sex-specific guidelines on stroke prevention7 could be com- tions, let alone hippocampus-linked cognitive functions, it is plemented by sex-specific guidelines on acute treatment deci- nonetheless notable that hippocampal regions, that were mark- sions, enhancing sex-sensitive stroke care and ultimately edly susceptible to hormonal influences as outlined above, con- increasing the benefit for both men and women. tributed to the female-specific lesion patterns identified in In this study, we explored sex disparities in lesion patterns of our study. global stroke severity, which allows for some broad clinical It is however important to note, that we did not have access to implications of our findings. Naturally, our results may to a information on the hormonal status in women. Instead, we here certain extent be dominated by effects due to motor symptoms, assumed the same median age of menopause for all female given their disproportionally high importance for the overall patients. Employing the actual age at menopause, as well as NIHSS score. To evaluate more specific brain functions at acute increasing the number of younger patients in general would have and chronic stages, future studies could therefore focus on sub- allowed for stronger conclusions on the potential organizational items of the NIHSS, such as language impairments, dysarthria, or activational character of findings. Besides, the availability of disturbances in orientation, neglect, or on specific cognitive the exact level of estrogens would be most desirable, as previous behavioral tests, e.g., probing memory functions. This would be a studies indicate that sex-specific functional cerebral asymmetries promising approach to trace back our most prominent sex- may even vary during the menstrual cycle27. Moreover, the specific finding, the relevance of the left PCA territory, to specific measurement of exact hormonal levels could inform about the brain functions. Incorporating outcomes from more chronic time effects of hormonotherapy (e.g., hormone replacement therapy points would furthermore allow for more definite conclusions on after menopause, oral contraceptives prior to menopause) and long-term effects. Above all, it may be especially fruitful to surgical interventions, such as hysterectomies, oophorectomies, examine whether sex differences observed here generalize to and ovaries remaining after hysterectomy. Therefore, future stu- cerebral reorganization processes during the recovery phase post- dies could attempt to gather data on hormonal status, and also to stroke. In the positive case, generated results could fuel sex- recruit more balanced numbers of men and women of pre- specific personalized clinical rehabilitation endeavors57. menopausal and postmenopausal status. In addition, it might be Similarly, the spatial resolution of our Bayesian hierarchical promising to test links between implicated brain regions and sex- approach stays within the realms of frequently arising, typical specific genetic underpinnings, as was recently introduced for lesion patterns, and back-transformed atlas brain regions. Con- healthy population samples47,48. sequently, our region-wise approach does not allow for a com- Independent of their exact origin, the sex-sensitive effects that parably high spatial resolution of lesion-symptom analyses we observed in brain-behavior associations underlying stroke relying on voxel-wise data18,34. Nonetheless, this reduction in severity, could lead to important clinical implications. Since lesion dimensionality was necessary to render sex-aware analyses lesions of any kind explained more severe stroke in women, within our Bayesian hierarchical model framework feasible. We rescuing the same (normalized) amount of brain tissue—for here employed two frequently used brain atlas templates58,59 and example, by thrombolysis or mechanical thrombectomy—could unsupervised dimensionality reduction22. Future studies could have a more enhanced effect in female than male patients. This explore differing atlas templates and also supervised strategies to resulting expectation is well in line with previous reports on organize brain regions. Such supervised approaches could reduce enhanced therapy response to intravenous thrombolysis49–51 or variability between lesion embeddings of different datasets, since more advantageous long-term outcome in women compared we here saw subtle differences in automatically derived lesion to men in clinical mechanical thrombectomy studies52. In atoms. Eventually, confidence in our results could then be this recently published study on sex discrepancies after increased in case of successful replication despite these metho- thrombectomy52, Sheth and colleagues ascertain these outcome dological alterations. What is more, our Bayesian models could, differences between men and women despite comparable infarct for example, be modified to integrate an additional hierarchical volumes and reperfusion rates. level to differentiate between rich-club and non-rich-club Most previous studies have relied on sex-independent, general regions60,61 or contrast agglomerated and separately modeled cutoffs of salvageable tissue volumes to decide, for example, sets of brain regions to test brain modes, as introduced by whether to undertake mechanical thrombectomy in the later time Godefroy and colleagues62,63. window53. A conceivable first step could be to revisit previous Lastly, we did carefully account for interindividual differences randomized clinical trials’ data to investigate rescued tissue–sex in important sociodemographic characteristics (e.g., age), interactions. Future clinical treatment studies could then pro- comorbidities (e.g., atrial fibrillation), markers of chronic brain spectively and systematically test varying cutoffs for men and health (e.g., white matter hyperintensity lesion load), and stroke women, hypothesizing that rescuing a lesser amount of tissue in subtypes (e.g., cardioembolic stroke genesis). We furthermore women would still be sufficient for a noticeable positive treatment adjusted all analyses for global, lesion atom-independent sex response. Furthermore, it may be important to take into account differences (e.g., due to a potentially longer time between symp- this female-specific salvaging effect for lesions in the posterior tom onset and hospital admission in women)10. In addition, each circulation territory of the left hemisphere. Thrombectomy for image and corresponding lesion mask was individually quality primary distal posterior cerebral artery occlusion stroke was controlled to ensure the absence of overt spatial distortions (e.g., recently found to be a safe and potentially beneficial treatment due to cerebral atrophies). However, our cohorts were slightly option, as treated patients showed a pronounced decrease in imbalanced with respect to the men:women ratio, which may not stroke severity until discharge54. In view of these results, future have faithfully captured the oftentimes reported equally high or thrombectomy studies could evaluate whether female patients higher stroke incidence in women64,65. Also, we did not have benefited even more substantially from these reperfusion thera- access to some measures, which could be of potential interest as pies of more distal PCA occlusions. Notably, such studies should covariate in our analysis, for a majority of patients; for example, also consider additional age–sex interactions, given that our delays in hospital arrival, the exact time between imaging and NATURE COMMUNICATIONS | (2021)12:3289 | https://doi.org/10.1038/s41467-021-23492-3 | www.nature.com/naturecommunications 9 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-23492-3 NIHSS acquisition, acute changes in stroke severity (e.g., due to of 62–116 ms, TI 2200 ms, FOV 220–240 mm). Ischemic DWI tissue lesions were spontaneous reperfusion, brain edema, or seizures) or adminis- manually outlined using semiautomated algorithms68. Raters were blinded to clinical outcomes. tered revascularization therapies and their interactions with stroke severity (e.g., NIHSS scores upon admission and after acute treatment may differ substantially). In this study, we aimed to Magnetic resonance imaging: preprocessing. Individual images were spatially address these factors indirectly by tailoring ancillary analyses to normalized to standard Montreal Neurological Institute (MNI-152) space: we first linearly realigned both DWI and FLAIR images with an MNI template. Subse- subgroups, e.g., only those patients whose data was acquired quently, we co-registered the DWI image to the FLAIR image, denoised the directly after admission. Their data acquisition thus preceded the images69, and lastly employed the unified segmentation algorithm to nonlinearly onsets of any acute treatment effects with high probability— normalize the FLAIR image70. We masked lesioned tissue during this normal- female-specific effects relating to left posterior circulation lesions ization step to mitigate the risk of image distortions71. The same transformation was applied to the DWI image, as well as the corresponding DWI-derived binary were nonetheless reliably observable. Future large-scale studies lesion mask. This preprocessing pipeline, especially featuring the co-registration of are warranted to recruit equal numbers of men and women and DWI and FLAIR images, was optimized to generate as reliable and accurate spatial systematically record further clinical aspects, such as delays in normalizations for as many patients as possible. The quality of normalized DWI- hospital arrival and acute treatments, to facilitate more compre- lesion masks was carefully inspected by two experienced raters (A.K.B. and M.B). Insufficient quality, predominantly arising from moderate to severe motion arti- hensive and explicit investigations. In particular, these studies facts and/or moderate to severe normalization errors led to the exclusion of may be enriched for patients receiving acute treatments to max- patients (c.f. Supplementary Information for details). The WMH volume (WMHv) imize the power of uncovering sex- and treatment-specific lesion was computed based on manual WMH outlines onto FLAIR images performed atom effects. By these means, it may eventually be possible to using the MRIcro software (04/2010, 08/2014, and 2019 versions)72. quantify the clinical importance of sex-specific lesion patterns by taking into account their modifying effects on acute recanaliza- Derivation of a low-dimensional lesion representation. We successively com- tion treatments. Since some evidence for example suggests sex- bined (1) an automated low-dimensional embedding of high-dimensional lesion specific effects of white matter integrity on stroke outcome66, information, and (2) probabilistic modeling to explain the acute stroke severity, as measured by the NIHSS (c.f. ref. 23, for a comparable analytical approach). While further future studies may also not only include acute, but also we initially determined eloquent lesion patterns across all patients, we subsequently chronic markers of brain health while taking sex into account refined analyses to investigate sex differences on different levels of our Bayesian more closely. hierarchical model. Women tend to have more severe strokes than men. Previous In view of the high-dimensional lesioned voxel space that would inevitably lead to severe overfitting in our regression analyses, we first captured the number of methodology did not allow to evaluate whether lesion patterns lesioned voxels within 109 brain regions, as defined by the Harvard-Oxford atlas58. contributed to these sex differences in outcomes. By deriving a More precisely, these brain regions represented 47 cortical and 7 subcortical gray low-dimensional lesion representation and employing Bayesian matter brain regions per hemisphere, as well as the brainstem. White matter hierarchical modeling, we here uncovered considerable sex dis- damage was read out in form of lesion load within each one of 20 John-Hopkins- University (JHU) atlas defined white matter tracts (7 hemisphere-specific tracts, 6 crepancies in lesion patterns explaining acute stroke severity. bilateral tracts)59,73. Thus, these steps resulted in 129 parcels in total. To extract While bilateral subcortical and left-hemispheric inferior frontal, reoccurring, archetypical and directly interpretable lesion patterns in our stroke superior, and middle temporal regions, i.e., presumed motor and population in a data-driven, unsupervised fashion, we performed NMF22 on the language regions, explained more severe strokes in both men and log-transformed brain region- and tract-wise lesion load. By these means, we obtained ten unique topographical lesion configurations, that we will call lesion women, effects in women were more widespread and similar atoms in the following (c.f., Supplementary Information for further details). lesions underlay more severe strokes in women compared to men. Probable vascular territories of lesion atoms were assigned via evaluations by In particular, lesions in the posterior circulation of the left individuals with several years of neurology and neuroradiology experience (A.K.B., hemisphere were associated with a higher stroke severity exclu- M.B., and N.R.). sively in women. These differences were robustly validated in a second independent, international multisite lesion dataset, and Explaining interindividual differences in acute stroke outcomes. These ten could not be explained by sex differences in lesion volume or NMF-derived lesion atoms, containing information on individual lesion patterns, more severe strokes in women in general. served as neuroimaging-derived input to our Bayesian hierarchical linear regres- sion model74 to explain acute stroke severity. We aimed to obtain fully probabilistic model parameter estimates that could inform us about each lesion atom’s influence Methods on the outcome. While we first computed the impact of lesion atoms across all Participant recruitment. AIS patients (n = 555), considered as derivation cohort patients in a first Bayesian model, we refined analyses in a subsequent step and in this study, were admitted to Massachusetts General Hospital and enrolled as of introduced a hierarchical lesion atom structure for a second model. This hierarchy part the Genes Associated with Stroke Risk and Outcomes Study (GASROS; mean allowed the stratification for an individual’s sex, i.e., we could estimate the lesion age(SD): 65.0(14.8) years, 38% females, c.f., Table 1 for sex-specific numbers)61,67. atom’s influence on the outcome in women and men separately. Hence, we Inclusion was generally considered for any AIS patient that met the following obtained separate lesion pattern relevances for men and women. criteria; (i) adult patients ≥18 years of age, (ii) admitted to the emergency Prior to carrying out the Bayesian model, lesion atom data, as well as the stroke department with signs and symptoms of AIS, and (iii) neuroimaging confirmation severity outcome score were corrected for lesion volume75. In addition, our model of an acute infarct. Only patients with MRI data obtained within 48 h from took into account the effects of (normalized) age, age2, sex, and the presence of the admission, as well as complete phenotypic data, such as stroke severity and stroke following comorbidities: hypertension, diabetes mellitus type 2, atrial fibrillation, risk factors were included in this study (i.e., complete case analyses, c.f., Supple- coronary artery disease, and lastly the log-transformed white matter hyperintensity mentary Information for a sample size derivation). Patients gave written informed lesion volume. In view of women’s overall higher age, we aimed to adjust for age- consent in accordance with the Declaration of Helsinki. The study protocol was specific effects particularly comprehensively by not only taking into account the approved by Massachusetts General Hospital’s Institutional Review Board (Pro- basic age, yet also its squared value and therefore U-shaped effects as well. tocol #: 2001P001186). Importantly, we also included sex as a nonhierarchical, lower-level variable in the model to capture sex differences in stroke severity that were independent of lesion patterns. Women may, for example, also present with more severe strokes on Stroke patient characteristics and imaging. Patients were examined by trained, average. They may be of a more advanced age and have a greater prestroke level of board-certified vascular neurologists. The recorded sociodemographic and clinical disability when experiencing the stroke, leading to a higher stroke severity, yet variables included age, sex, and common vascular risk factors (hypertension, dia- decreased likelihood of acute recanalization treatment. Such sex-specific, but lesion betes mellitus type 2, atrial fibrillation, coronary artery disease). Stroke severity was pattern-independent differences in stroke severity could then be modeled by the captured by the acute NIHSS (0: no symptoms, 42: maximum stroke severity). sex variable. It would, however, not have an influence on lesion pattern–sex Each patient was scanned within 48 h of admission, standardized clinical interaction effects. In other words, the nonhierarchical sex variable represents the imaging protocols included DWI (in the majority of cases on 1.5 T General Electric “residual” sex effects on the outcome that are not already captured by the Signa scanners, and a few cases on 1.5 or 3 T Siemens scanners, repetition time hierarchical sex-specific NMF components. (TR) 5000 ms, minimum echo time (TE) of 62–117 ms, field-of-view (FOV) 220 The full Bayesian hierarchical model specification was as follows: mm field-of-view, 5-mm slice thickness with a 1-mm gap, and 0 s/mm2 (b-zero) Hyperpriors and 1000 s/mm2 b-values), and axial T2 FLAIR images (TR 5000 ms, minimum TE hyper_σ_ β ∼ Halfcauchy(5) 10 NATURE COMMUNICATIONS | (2021)12:3289 | https://doi.org/10.1038/s41467-021-23492-3 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-23492-3 ARTICLE σ_ βm,f ∼ Halfcauchy(hyper_ σ_ β) We then contrasted male- and female-specific lesion pattern effects within the hyper_mu_ β ∼ Normal(µ = 0, σ = 10) cardioembolic and non-cardioembolic stroke subgroups. To allow for a reasonably mu_ βm,f ∼ Normal(µ = hyper_ mu_ β, σ = 10) high number of patients in each of these four subgroups, we here merged data of Priors the derivation and validation cohorts (c.f., Supplementary Information for further α ∼ Normal(µ = 0, σ = 1) details). β1-10; m,f ∼ Normal(µ = mu_ βm,f, σ = σ_ βm,f) Lastly, we aimed to explore whether possible sex differences were more likely to βage ∼ Normal(µ =0, σ = 10) be of organizational or activational hormonal nature. Since we did not have access βage*age ∼ Normal(µ = 0, σ = 10) to the precise hormonal status in women, we stratified our sample based on an age βsex ∼ Normal(µ = 0, σ = 1) cutoff of 52 years, according to the median age at menopause26. As in the ancillary βhypertension ∼ Normal(µ = 0, σ = 1) analysis focused on stroke subtype, we inserted an additional hierarchical level and βdiabetes ∼ Normal(µ = 0, σ = 1) build two groups of patients above (≥) 52 and below (<52) years of age within βatrial fibrillation ∼ Normal(µ =0, σ = 1) which we compared lesion pattern effects of men and women. We once again βcoronary artery disease ∼ Normal(µ = 0, σ = 1) merged derivation and validation cohort data to maximize the number of available βWMHv ∼ Normal(µ = 0, σ = 1) patients, especially in the subgroup of younger stroke patients. Likelihood of linear model NIHSS_est = α + β1–10[sex]* NMF-Component1-10 + βage* Age+βage*age* Age2 Reporting summary. Further information on research design is available in the Nature + βsex* Sex + βhypertension * hypertension + βdiabetes * diabetes + βatrial fibrillation * Research Reporting Summary linked to this article. atrial fibrillation + βcoronary artery disease * coronary artery disease + βWMHv * WMHv Model likelihood eps ∼ Halfcauchy(20) Data availability stroke_severity ∼ Normal(µ = NIHSS_est, σ = eps) The authors agree to make the data of development and validation cohorts available to We employed the No U-Turn Sampler (NUTS), a type of Monte Carlo Markov any researcher for the express purposes of reproducing the here presented results, and Chain algorithm (setting: draws = 5000)76, to draw samples from the Bayesian with the explicit permission for data sharing by Massachusetts General Hospital’s posterior parameter distributions. institutional review board. The Harvard-Oxford and JHU DTI-based white matter atlases Main advantages of our analytical approach may be seen in (i) the enhanced are accessible online (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/Atlases). Source data are interpretability of estimated NMF components, i.e., lesion atoms, (ii) the provided with this paper. multivariate nature of the stroke lesion embedding, and (iii) the possibility of integrating interaction effects between lesion atoms and an individual’s sex in a statistically sound probabilistic modeling framework. We thus make a step towards Code availability providing an alternative approach to classical VLSM analyses18, that estimate the Preprocessing of MRI images was conducted in a Matlab 2019b framework (The Mathworks, magnitude of effects in one brain location at a time and thus fit one model per Natick, MA, USA), the packages Statistical Parametric Mapping (SPM12; http://www.fil.ion. voxel. Our approach opens an alternative window to “one voxel at a time” by fitting ucl.ac.uk/spm/), and the ancillary package SPM_Superres (https://github.com/brudfors/ a single model to the entire brain. All ten of our lesion atoms were entered into the spm_superres). Further analyses were implemented in Python 3.7 (primarily packages: regression model at once. Hence, we could estimate the effect of each lesion atom, nilearn84 and pymc3; ref. 85). Full code for reproducibility and reuse is available here: https:// while accounting for the effects of all further lesion atoms. By these means, our github.com/AnnaBonkhoff/BMH_stroke_severity_sex_differences86. approach resembles more machine-learning-based multivariate lesion-symptom mapping methods, that have been introduced recently to evaluate lesion patterns instead of single voxels in association to an outcome77. Examples can be seen in Received: 8 October 2020; Accepted: 30 April 2021; approaches leveraging support vector machines34,78,79, tree-based80, or game theory-based algorithms81. Importantly, all of these multivariate approaches may decrease the distortion of functional localization compared to univariate voxel-wise approaches by considering subcortically and cortically located lesions at once20,82. We additionally aimed to mitigate confounding effects of excessively correlated neighboring voxels, as naturally arising due to the stroke lesion-vasculature References dependence, by explicitly combining frequently jointly affected brain regions in 1. Feigin, V. L. et al. Global and regional burden of stroke during 1990–2010: lesion patterns, i.e., our lesion atoms. Despite these methodological innovations, no findings from the Global Burden of Disease Study 2010. Lancet 383, 245–255 currently existing stroke modeling approach may achieve unfailing causal (2014). conclusions20. 2. Kelly-Hayes, M. et al. The influence of gender and age on disability following ischemic stroke: the Framingham study. J. Stroke Cerebrovasc. Dis. 12, Validation analysis. We repeated analyses in an independent dataset of 503 119–126 (2003). ischemic stroke patients (mean age(SD): 65.0(14.6), sex: 40.6% female, mean 3. Benjamin, E. J. et al. Heart disease and stroke statistics-2017 update: a report NIHSS(SD): 5.48(5.35), c.f., Supplementary Table 1 for further clinical character- from the American Heart Association. Circulation 135, e146–e603 (2017). istics), acquired within the framework of the multisite MRI-GENIE study24, to test 4. Bushnell, C. et al. Sex differences in the evaluation and treatment of acute the robustness of our findings (c.f., Supplementary Information for details). To ischaemic stroke. Lancet Neurol. 17, 641–650 (2018). further ascertain the robustness of our unsupervised lesion embedding, we corre- 5. Carcel, C., Woodward, M., Wang, X., Bushnell, C. & Sandset, E. C. Sex matters lated NMF components as computed for the derivation and validation cohort. in stroke: a review of recent evidence on the differences between women and Moreover, we recomputed region-wise relevances in final nonhierarchical analyses men. Front. Neuroendocrinol. 59, 100870 (2020). based on the other cohort’s lesion embedding (i.e., we first estimated the NMF 6. Giroud, M. et al. Temporal trends in sex differences with regard to stroke transformation based on one cohort’s imaging data and then applied this trans- incidence: the dijon stroke registry (1987–2012). Stroke 48, 846–849 (2017). formation to the data of the other cohort). We then ran additional correlation 7. Bushnell, C. et al. Guidelines for the prevention of stroke in women: a analyses between region-wise relevances as originating from the cohort-specific statement for healthcare professionals from the American Heart Association/ embedding, as well as the other cohort’s embedding to ensure our results were not American Stroke Association. Stroke 45, 1545–1588 (2014). substantially altered by the particular embedding. 8. Lisabeth, L. D., Brown, D. L., Hughes, R., Majersik, J. J. & Morgenstern, L. B. Acute stroke symptoms: comparing women and men. Stroke 40, 2031–2036 (2009). Ancillary analyses. We performed three ancillary hierarchical analyses. Firstly, we 9. Jerath, N. U., Reddy, C., Freeman, W. D., Jerath, A. U. & Brown, R. D. Gender aimed to reduce potential confounds arising from dispersed time points of data differences in presenting signs and symptoms of acute ischemic stroke: a acquisition and effects of acute recanalization therapies and limited the derivation population-based study. Gend. Med. 8, 312–319 (2011). cohort sample to only those patients whose data was acquired directly upon 10. Mandelzweig, L., Goldbourt, U., Boyko, V. & Tanne, D. Perceptual, social, and admission (and not within the first 48 h, as for the entire cohort). Thus, their data behavioral factors associated with delays in seeking medical care in patients acquisition likely preceded the potential administration of any acute recanalization with symptoms of acute stroke. Stroke 37, 1248–1253 (2006). therapy and their lesion-symptom associations may be considered to be purer. 11. Beal, C. C. Women’s interpretation of and cognitive and behavioral responses Conceivably, female-specific lesion effects could be due to a higher frequency of to the symptoms of acute ischemic stroke. J. Neurosci. Nurs. 46, 256–266 cardioembolic strokes and associated typical lesion patterns (e.g., more multifocal (2014). stroke lesions)83. In a second ancillary analysis, we therefore sought to disentangle 12. Friberg, L., Benson, L., Rosenqvist, M. & Lip, G. Y. H. Assessment of female stroke-subtype- and sex-specific effects, and concurrently performed stroke- sex as a risk factor in atrial fibrillation in Sweden: nationwide retrospective subtype- and sex-aware analyses, featuring two hierarchical levels. Accordingly, cohort study. BMJ 344, e3522–e3522 (2012). patients were not only stratified into groups of men and women as in main 13. Lang, C. et al. Do women with atrial fibrillation experience more severe analyses, yet initially assigned to a group of either cardioembolic or non- strokes?: results from the austrian stroke unit registry. Stroke 48, 778–780 cardioembolic stroke genesis, and then split into men and women (c.f., (2017). Supplementary Information for a comprehensive display of model specifications). NATURE COMMUNICATIONS | (2021)12:3289 | https://doi.org/10.1038/s41467-021-23492-3 | www.nature.com/naturecommunications 11 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-23492-3 14. Dehlendorff, C., Andersen, K. K. & Olsen, T. S. Sex disparities in stroke: 45. Hara, Y. et al. Synaptic correlates of memory and menopause in the women have more severe strokes but better survival than men. J. Am. Heart hippocampal dentate gyrus in rhesus monkeys. Neurobiol. Aging 33, 421. Assoc. 4, e001967 (2015). e17–421.e28 (2012). 15. Bonkhoff, A. K., Karch, A., Weber, R., Wellmann, J. & Berger, K. Female 46. Jacobs, E. G. et al. Reorganization of functional networks in verbal working stroke: sex differences in acute treatment and early outcomes of acute ischemic memory circuitry in early midlife: the impact of sex and menopausal status. stroke. Stroke 52, 406–415 (2021). Cereb. Cortex 27, 2857–2870 (2016). 16. Silva, G. S. et al. Gender differences in outcomes after ischemic stroke: role of 47. Liu, S., Seidlitz, J., Blumenthal, J. D., Clasen, L. S. & Raznahan, A. Integrative ischemic lesion volume and intracranial large-artery occlusion. Cerebrovasc. structural, functional, and transcriptomic analyses of sex-biased brain Dis. 30, 470–475 (2010). organization in humans. Proc. Natl Acad. Sci. USA 117, 18788–18798 (2020). 17. Hier, D. B., Yoon, W. B., Mohr, J. P., Price, T. R. & Wolf, P. A. Gender and 48. Smith, S. M. et al. An expanded set of genome-wide association studies of aphasia in the stroke data bank. Brain Lang. 47, 155–167 (1994). brain imaging phenotypes in UK Biobank. Nature Neuroscience. 24, 737–745 18. Bates, E. et al. Voxel-based lesion–symptom mapping. Nat. Neurosci. 6, 448 (2021) (2003). 49. Kent, D. M., Price, L. L., Ringleb, P., Hill, M. D. & Selker, H. P. Sex-based 19. Karnath, H.-O. The anatomy of spatial neglect based on voxelwise statistical differences in response to recombinant tissue plasminogen activator in acute analysis: a study of 140 patients. Cereb. Cortex 14, 1164–1172 (2004). ischemic stroke: a pooled analysis of randomized clinical trials. Stroke 36, 20. Sperber, C. Rethinking causality and data complexity in brain lesion- 62–65 (2005). behaviour inference and its implications for lesion-behaviour modelling. 50. Shobha, N., Sylaja, P. N. & Kapral, M. K. Differences in stroke outcome based Cortex 126, 49–62 (2020). on sex. Neurology 74, 767–771 (2010). 21. Wu, O. et al. Role of acute lesion topography in initial ischemic stroke severity 51. Lee, S.-J., Heo, S. H., Ambrosius, W. T. & Bushnell, C. D. Factors mediating and long-term functional outcomes. Stroke 46, 2438–2444 (2015). outcome after stroke: gender, thrombolysis, and their interaction. Transl. 22. Lee, D. D. & Seung, H. S. Learning the parts of objects by non-negative matrix Stroke Res. 9, 267–273 (2018). factorization. Nature 401, 788 (1999). 52. Sheth, S. A. et al. Sex differences in outcome after endovascular stroke therapy 23. Bonkhoff, A. K. et al. Generative lesion pattern decomposition of cognitive for acute ischemic stroke. Stroke 50, 2420–2427 (2019). impairment after stroke. Preprint at bio. Rxiv. https://doi.org/10.1101/ 53. Albers, G. W. et al. Thrombectomy for stroke at 6 to 16 hours with selection 2020.08.22.262873 (2020). by perfusion imaging. N. Engl. J. Med. 378, 708–718 (2018). 24. Giese, A.-K. et al. Design and rationale for examining neuroimaging genetics 54. Meyer, L. et al. Thrombectomy for primary distal posterior cerebral artery in ischemic stroke: the MRI-GENIE study. Neurol. Genet. 3, e180 (2017). occlusion stroke: the TOPMOST Study. JAMA Neurol. 78, 434–444 (2021). 25. Koellhoffer, E. C. & McCullough, L. D. The effects of estrogen in ischemic 55. Hinman, J. D. et al. Principles of precision medicine in stroke. J. Neurol. stroke. Transl. Stroke Res. 4, 390–401 (2013). Neurosurg. Psychiatry 88, 54–61 (2017). 26. McKinlay, S. M., Brambilla, D. J. & Posner, J. G. The normal menopause 56. Mauvais-Jarvis, F. et al. Sex and gender: modifiers of health, disease, and transition. Maturitas 14, 103–115 (1992). medicine. Lancet 396, 565–582 (2020). 27. Hausmann, M. Why sex hormones matter for neuroscience: a very short 57. Nachev, P., Rees, G. & Frackowiak, R. Lost in translation. F1000Research 7, review on sex, sex hormones, and functional brain asymmetries. J. Neurosci. 620 (2018). Res. 95, 40–49 (2017). 58. Desikan, R. S. et al. An automated labeling system for subdividing the human 28. Hirnstein, M., Hugdahl, K. & Hausmann, M. Cognitive sex differences and cerebral cortex on MRI scans into gyral based regions of interest. Neuroimage hemispheric asymmetry: a critical review of 40 years of research. Laterality 24, 31, 968–980 (2006). 204–252 (2019). 59. Mori, S., Wakana, S., Van Zijl, P. C. & Nagae-Poetscher, L. M. MRI Atlas of 29. McGlone, J. Sex differences in human brain asymmetry: a critical survey. Human White Matter (Elsevier, 2005). Behav. Brain Sci. 3, 215–227 (1980). 60. Van Den Heuvel, M. P. & Sporns, O. Rich-club organization of the human 30. Inglis, J., Ruckman, M., Lawson, J. S., MacLean, A. W. & Monga, T. N. Sex connectome. J. Neurosci. 31, 15775–15786 (2011). differences in the cognitive effects of unilateral brain damage. Cortex 18, 61. Schirmer, M. D. et al. Rich-club organization: an important determinant of 257–275 (1982). functional outcome after acute ischemic stroke. Front. Neurol. 10, 956 (2019). 31. Inglis, J. & Lawson, J. S. A meta-analysis of sex differences in the effects of 62. Godefroy, O. et al. Brain-behaviour relationships. Some models and related unilateral brain damage on intelligence test results. Can. J. Psychol. 36, 670 statistical procedures for the study of brain-damaged patients. Brain: a J. (1982). Neurol. 121, 1545–1556 (1998). 32. Inglis, J., Ruckman, M., Lawson, J. S., MacLean, A. W. & Monga, T. N. Sex 63. Toba, M. N. et al. Revisiting ‘Brain Modes’ in a New Computational Era: differences in the cognitive effects of unilateral brain damage: comparison of Approaches for the Characterization of Brain-behavioural Associations (Oxford stroke patients and normal control subjects. Cortex 19, 551–555 (1983). University Press, 2020). 33. Cheng, B. et al. Influence of stroke infarct location on functional outcome 64. Madsen, T. E. Sex-specific stroke incidence over time in the Greater measured by the modified rankin scale. Stroke 45, 1695–1702 (2014). Cincinnati/Northern Kentucky Stroke Study. Neurology 89, 990–996 (2017). 34. Smith, D. V., Clithero, J. A., Rorden, C. & Karnath, H.-O. Decoding the 65. Kim, J. et al. Global stroke statistics 2019. Int. J. Stroke 15, 819–838 (2020). anatomical network of spatial attention. Proc. Natl Acad. Sci. USA 110, 66. Etherton, M. R. et al. Sex-specific differences in white matter microvascular 1518–1523 (2013). integrity after ischaemic stroke. Stroke Vasc. Neurol. 4, 198–205 (2019). 35. Siegel, J. S. et al. Disruptions of network connectivity predict impairment in 67. Schirmer, M. D. et al. Spatial signature of white matter hyperintensities in multiple behavioral domains after stroke. Proc. Natl Acad. Sci. USA 113, stroke patients. Front. Neurol. 10, 208–208 (2019). E4367–E4376 (2016). 68. Rorden, C. & Brett, M. Stereotaxic display of brain lesions. Behav. Neurol. 12, 36. Reeves, M. J. et al. Sex differences in stroke: epidemiology, clinical 191–200 (2000). presentation, medical care, and outcomes. Lancet Neurol. 7, 915–926 (2008). 69. Brudfors, M., Balbastre, Y., Nachev, P. & Ashburner, J. A Tool for super- 37. McCarthy, M. M. & Arnold, A. P. Reframing sexual differentiation of the resolving multimodal clinical MRI. Preprint at https://arxiv.org/abs/ brain. Nat. Neurosci. 14, 677–683 (2011). 1909.01140 (2019). 38. Hurn, P. D. & Macrae, I. M. Estrogen as a neuroprotectant in stroke. J. Cereb. 70. Ashburner, J. & Friston, K. J. Unified segmentation. Neuroimage 26, 839–851 Blood Flow. Metab. 20, 631–652 (2000). (2005). 39. Viscoli, C. M. et al. A clinical trial of estrogen-replacement therapy after 71. Brett, M., Leff, A. P., Rorden, C. & Ashburner, J. Spatial normalization of brain ischemic stroke. N. Engl. J. Med. 345, 1243–1249 (2001). images with focal lesions using cost function masking. Neuroimage 14, 40. Nudy, M., Chinchilli, V. M. & Foy, A. J. A systematic review and meta- 486–500 (2001). regression analysis to examine the ‘timing hypothesis’ of hormone 72. Rost, N. S. et al. White matter hyperintensity volume is increased in small replacement therapy on mortality, coronary heart disease, and stroke. Int. J. vessel stroke subtypes. Neurology 75, 1670–1677 (2010). Cardiol. Heart Vasc. 22, 123–131 (2019). 73. Wakana, S. et al. Reproducibility of quantitative tractography methods applied 41. Alkayed, N. J. et al. Gender-linked brain injury in experimental stroke. Stroke to cerebral white matter. Neuroimage 36, 630–644 (2007). 29, 159–165 (1998). 74. Gelman, A. & Hill, J. Data Analysis Using Regression and Multilevel/ 42. McCullough, L. D., Alkayed, N. J., Traystman, R. J., Williams, M. J. & Hurn, P. hierarchical Models (Cambridge University Press, 2006). D. Postischemic estrogen reduces hypoperfusion and secondary ischemia after 75. Schwartz, M. F., Faseyitan, O., Kim, J. & Coslett, H. B. The dorsal stream experimental stroke. Stroke 32, 796–802 (2001). contribution to phonological retrieval in object naming. Brain 135, 3799–3814 43. Liu, M., Dziennis, S., Hurn, P. D. & Alkayed, N. J. Mechanisms of gender- (2012). linked ischemic brain injury. Restor. Neurol. Neurosci. 27, 163–179 (2009). 76. Hoffman, M. D. & Gelman, A. The No-U-Turn sampler: adaptively setting 44. Gibson, C. L. Cerebral ischemic stroke: is gender important? J. Cereb. Blood path lengths in Hamiltonian Monte Carlo. J. Mach. Learn. Res. 15, 1593–1623 Flow. Metab. 33, 1355–1361 (2013). (2014). 12 NATURE COMMUNICATIONS | (2021)12:3289 | https://doi.org/10.1038/s41467-021-23492-3 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-23492-3 ARTICLE 77. Vaidya, A. R., Pujara, M. S., Petrides, M., Murray, E. A. & Fellows, L. K. Lesion Author contributions studies in contemporary neuroscience. Trends Cogn. Sci. 23, 653–671 (2019). A.K.B., D.B., and N.S.R. conceived and designed the study, led data interpretation, and 78. Zhang, Y., Kimberg, D. Y., Coslett, H. B., Schwartz, M. F. & Wang, Z. prepared the manuscript. A.K.B. led data analysis, M.D.S., M.B., S.H., R.W.R., M.B., and Multivariate lesion‐symptom mapping using support vector regression. Hum. O.W. also contributed to data analysis. E.D. led data management, data preprocessing, Brain Mapp. 35, 5861–5876 (2014). and execution of data analysis workflows. K.L.D., M.J.N., A.V.D., A.K.G., M.R.E., B.L.H., 79. Yourganov, G., Fridriksson, J., Rorden, C., Gleichgerrcht, E. & Bonilha, L. B.J.T.M., E.C.M., J.A., O.S.B., S.B., J.W.C., A.D., C.J.G., La.H., Lu.H., K.J., J.J.C., S.J.K., Multivariate connectome-based symptom mapping in post-stroke patients: R.L., C.R.L., C.W.M., J.F.M., C.L.P., A.R., S.R., Jo.R., Ja.R., T.R., R.L.S., R.S., P.S., Ag.S., networks supporting language and speech. J. Neurosci. 36, 6668–6679 (2016). M.S., Al.S., T.M.S, D.S., T.T., V.T., A.V., J.W., D.W., R.Z., P.F.M., B.B.W., C.J., A.G.L., 80. Pustina, D. et al. Enhanced estimations of post-stroke aphasia severity using J.M., and O.W. contributed to data acquisition, management, and preprocessing. All stacked multimodal predictions: enhanced predictions of aphasia severity. authors contributed to results interpretation and final manuscript preparation. Hum. Brain Mapp. 38, 5603–5615 (2017). 81. Toba, M. N. et al. Game theoretical mapping of causal interactions underlying visuo-spatial attention in the human brain based on stroke lesions. Hum. Competing interests M.E. has received personal fees for consulting from Astra Zeneca and WorldCare Clinical Brain Mapp. 38, 3454–3471 (2017). Group. C.G. has received consulting honoraria from Microvention and Strykere, and 82. Mah, Y.-H., Husain, M., Rees, G. & Nachev, P. Human brain lesion-deficit research funding from Medtronic and Penumbra. A.V. has received research funding inference remapped. Brain 137, 2522–2531 (2014). from Cerenovus. A.G.L. has received personal fees from Bayer, Astra Zeneca, BMS Pfizer, 83. Sharobeam, A. et al. Patterns of infarction on MRI in patients with acute and Portola. N.S.R. has received compensation as scientific advisory consultant from ischemic stroke and cardio-embolism: a systematic review and meta-analysis. Omniox, Sanofi Genzyme, and AbbVie Inc. All other authors declare no competing Front. Neurol. 11, 1699 (2020). interests. 84. Abraham, A. et al. Machine learning for neuroimaging with scikit-learn. Front. Neuroinformatics 8, 14 (2014). 85. Salvatier, J., Wiecki, T. V. & Fonnesbeck, C. Probabilistic programming in Additional information Python using PyMC3. PeerJ Computer Sci. 2, e55 (2016). Supplementary information The online version contains supplementary material 86. Bonkhoff AK et al. Outcome after acute ischemic stroke is linked to sex-specific available at https://doi.org/10.1038/s41467-021-23492-3. lesion patterns. Zenodo. https://doi.org/10.5281/zenodo.4670275 (2021). 87. Carcel, C. et al. Trends in recruitment of women and reporting of sex Correspondence and requests for materials should be addressed to A.K.B. differences in large-scale published randomized controlled trials in stroke. Int. J. Stroke 14, 931–938 (2019). Peer review information Nature Communications thanks Cheryl Bushnell, Argye Hillis, Monica Toba, and the other, anonymous reviewer(s) for their contribution to the peer review of this work. Peer review reports are available. Acknowledgements We are grateful to our colleagues at the J. Philip Kistler Stroke Research Center for Reprints and permission information is available at http://www.nature.com/reprints valuable support and discussions and would like to specifically acknowledge valuable comments by Parashkev Nachev, M.D., Ph.D. on previous versions on this manuscript. Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in Furthermore, we are grateful to our research participants without whom this work would published maps and institutional affiliations. not have been possible. M.B. acknowledges support from the Société Française de Neuroradiologie, Société Française de Radiologie, and Fondation ISITE-ULNE. A.V. is in part supported by NIH-NINDS (R01 NS103824, RF1 NS117643, R01 NS100417, U01NS100699, and U01NS110772). C.J. acknowledges support from the Swedish Open Access This article is licensed under a Creative Commons Research Council (2018-02543), the Swedish state under the agreement between the Attribution 4.0 International License, which permits use, sharing, Swedish government and the county councils, the ALF agreement (ALFGBG-720081); adaptation, distribution and reproduction in any medium or format, as long as you give the Swedish Heart and Lung Foundation (20190203). A.G.L. acknowledges support from appropriate credit to the original author(s) and the source, provide a link to the Creative the Swedish Research Council (2019-01757), The Swedish Government (under the “Avtal Commons license, and indicate if changes were made. The images or other third party om Läkarutbildning och Medicinsk Forskning, ALF”), The Swedish Heart and Lung material in this article are included in the article’s Creative Commons license, unless Foundation, Region Skåne, Lund University, Skåne University Hospital, Sparbanks- indicated otherwise in a credit line to the material. If material is not included in the stiftelsen Färs och Frosta, Fremasons Lodge of Instruction Eos in Lund and NIH article’s Creative Commons license and your intended use is not permitted by statutory (1R01NS114045-01). D.B. has been funded by the Brain Canada Foundation, through the regulation or exceeds the permitted use, you will need to obtain permission directly from Canada Brain Research Fund, with the financial support of Health Canada, National the copyright holder. To view a copy of this license, visit http://creativecommons.org/ Institutes of Health (NIH R01 AG068563A), the Canadian Institute of Health Research licenses/by/4.0/. (CIHR 438531), the Healthy Brains Healthy Lives initiative (Canada First Research Excellence fund), Google (Research Award, Teaching Award), and by the CIFAR Arti- ficial Intelligence Chairs program (Canada Institute for Advanced Research). N.S.R. is in © The Author(s) 2021 part supported by NIH-NINDS (R01NS082285, R01NS086905, and U19NS115388). 1 J. Philip Kistler Stroke Research Center, Massachusetts General Hospital, Harvard Medical School, Boston, MA, USA. 2Clinic for Neuroradiology, University Hospital Bonn, Bonn, Germany. 3Univ. Lille, Inserm, CHU Lille, U1171 – LilNCog (JPARC) – Lille Neurosciences & Cognition, F-59000, Lille, France. 4School of Biomedical Engineering & Imaging Sciences, King’s College London, London, UK. 5Computer Science and Artificial Intelligence Lab, Massachusetts Institute of Technology, Boston, MA, USA. 6Athinoula A. Martinos Center for Biomedical Imaging, Department of Radiology, Massachusetts General Hospital, Charlestown, MA, USA. 7Department of Neurology, University Medical Center Hamburg‐Eppendorf, Hamburg, Germany. 8Hunter Medical Research Institute, Newcastle, NSW, Australia. 9School of Medicine and Public Health, University of Newcastle, Newcastle, NSW, Australia. 10Department of Medicine, Division of Neurology, University of British Columbia, Vancouver, BC, Canada. 11 School of Life Sciences, University of Lincoln, Lincoln, UK. 12Department of Neurology, University of Maryland School of Medicine and Veterans Affairs Maryland Health Care System, Baltimore, MD, USA. 13School of Medical Sciences, University of Campinas (UNICAMP) and the Brazilian Institute of Neuroscience and Neurotechnology (BRAINN), Campinas, Sao Paulo, Brazil. 14Department of Neurosurgery, Geisinger, Danville, PA, USA. 15Research Institute of Neurointervention, Paracelsus Medical University, Salzburg, Austria. 16Department of Emergency Medicine, Washington University School of Medicine, St Louis, MO, USA. 17Department of Neurology, Washington University School of Medicine & Barnes- Jewish Hospital, St Louis, MO, USA. 18Department of Clinical Neuroscience, Institute of Neuroscience and Physiology, Sahlgrenska Academy, University of Gothenburg, Gothenburg, Sweden. 19Department of Neurology, Sahlgrenska University Hospital, Gothenburg, Sweden. 20Department of Neurology, Neurovascular Research Group (NEUVAS), IMIM-Hospital del Mar (Institut Hospital del Mar d’Investigacions Mèdiques), Universitat Autonoma de Barcelona, Barcelona, Spain. 21KU Leuven - University of Leuven, Department of Neurosciences, Experimental Neurology and Leuven Research Institute for Neuroscience and Disease (LIND), Leuven, Belgium. 22VIB, Vesalius Research Center, Laboratory of Neurobiology, University Hospitals Leuven, Department of Neurology, Leuven, Belgium. 23School of Medicine and Public Health, University of NATURE COMMUNICATIONS | (2021)12:3289 | https://doi.org/10.1038/s41467-021-23492-3 | www.nature.com/naturecommunications 13 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-021-23492-3 Newcastle, Newcastle, NSW, Australia. 24Department of Neurology, John Hunter Hospital, Newcastle, NSW, Australia. 25Department of Pharmacotherapy and Translational Research and Center for Pharmacogenomics, University of Florida, Gainesville, FL, USA. 26Department of Neurology, Mayo Clinic, Jacksonville, FL, USA. 27Centogene AG, Rostock, Germany. 28Department of Neurology, Clinical Division of Neurogeriatrics, Medical University Graz, Graz, Austria. 29Center for Genomic Medicine, Massachusetts General Hospital, Boston, MA, USA. 30 Department of Neurology and Evelyn F. McKnight Brain Institute, Miller School of Medicine, University of Miami, Miami, FL, USA. 31Institute of Cardiovascular Research, Royal Holloway University of London (ICR2UL), Egham, UK. 32St Peter’s and Ashford Hospitals, Egham, UK. 33 Department of Neurology, Jagiellonian University Medical College, Krakow, Poland. 34Department of clinical sciences Malmö, Lund University, Lund, Sweden. 35Department of Neurology, Skåne University Hospital, Lund and Malmö, Sweden. 36Department of Laboratory Medicine, Institute of Biomedicine, the Sahlgrenska Academy at University of Gothenburg, Gothenburg, Sweden. 37Department of Neurology, Helsinki University Hospital and University of Helsinki, Helsinki, Finland. 38Department of Clinical Neuroscience, Institute of Neuroscience and Physiology, Sahlgrenska Academy at University of Gothenburg, Gothenburg, Sweden. 39Department of Neurology, Sahlgrenska University Hospital, Gothenburg, Sweden. 40 Stroke Division, Florey Institute of Neuroscience and Mental Health, Heidelberg, Australia. 41Department of Neurology, Austin Health, Heidelberg, Australia. 42Department of Radiology, University of Cincinnati College of Medicine, Cincinnati, OH, USA. 43Department of Clinical Sciences Lund, Radiology, Lund University, Lund, Sweden. 44Department of Radiology, Neuroradiology, Skåne University Hospital, Lund, Sweden. 45 Department of Neurology and Rehabilitation Medicine, University of Cincinnati College of Medicine, Cincinnati, OH, USA. 46Department of Neurology, Geisinger, Danville, PA, USA. 47Division of Endocrinology, Diabetes and Nutrition, Department of Medicine, University of Maryland School of Medicine, Baltimore, MD, USA. 48Departments of Neurology and Public Health Sciences, University of Virginia, Charlottesville, VA, USA. 49 Department of Clinical Genetics and Genomics, Sahlgrenska University Hospital, Gothenburg, Sweden. 50Department of Neurology, Skåne University Hospital, Lund, Sweden. 51Department of Clinical Sciences Lund, Neurology, Lund University, Lund, Sweden. 52University of Technology Sydney, Sydney, NSW, Australia. 53Department of Biomedical Engineering, McConnell Brain Imaging Centre, Montreal Neurological Institute, Faculty of Medicine, School of Computer Science, McGill University, Montreal, QC, Canada. 54Mila - Quebec Artificial Intelligence Institute, Montreal, QC, Canada. *A list of authors and their affiliations appears at the end of the paper. ✉email: abonkhoff@mgh.harvard.edu MRI-GENIE and GISCOME Investigators and the International Stroke Genetics Consortium Anna K. Bonkhoff 1 ✉, Markus D. Schirmer1,2, Martin Bretzner 1,3, Sungmin Hong1, Robert W. Regenhardt 1, Mikael Brudfors4, Kathleen L. Donahue1, Marco J. Nardin1, Adrian V. Dalca5,6, Anne-Katrin Giese7, Mark R. Etherton1, Brandon L. Hancock6, Steven J. T. Mocking6, Elissa C. McIntosh6, John Attia 8,9, Oscar R. Benavente10, Stephen Bevan11, John W. Cole12, Amanda Donatti13, Christoph J. Griessenauer14,15, Laura Heitsch16,17, Lukas Holmegaard18,19, Katarina Jood18,19, Jordi Jimenez-Conde20, Steven J. Kittner12, Robin Lemmens21,22, Christopher R. Levi23,24, Caitrin W. McDonough25, James F. Meschia 26, Chia-Ling Phuah17, Arndt Rolfs27, Stefan Ropele 28, Jonathan Rosand1,6,29, Jaume Roquer20, Tatjana Rundek30, Ralph L. Sacco30, Reinhold Schmidt28, Pankaj Sharma31,32, Agnieszka Slowik33, Martin Söderholm34,35, Alessandro Sousa13, Tara M. Stanne36, Daniel Strbian 37, Turgut Tatlisumak38,39, Vincent Thijs40,41, Achala Vagal42, Johan Wasselius43,44, Daniel Woo45, Ramin Zand46, Patrick F. McArdle47, Bradford B. Worrall 48, Christina Jern36,49, Arne G. Lindgren50,51, Jane Maguire52, Danilo Bzdok 53,54, Ona Wu6 & Natalia S. Rost1 14 NATURE COMMUNICATIONS | (2021)12:3289 | https://doi.org/10.1038/s41467-021-23492-3 | www.nature.com/naturecommunications