Risk and protective factors for stroke and transient ischaemic attack in patients with atrial fibrillation: an umbrella review of meta-analyses Authors: Axel Kylhammar, Dimitrios Tsartsalis, Ziad Hijazi, Constantina Aggeli, Costas Tsioufis, Lars O Karlsson, Dimitrios Venetsanos, Frieder Braunschweig, Nikola Drca, Astrid Paul-Nordin, Finn Akerström, Emma Svennberg, Emmanouil Charitakis Published: 2026 ABSTRACT AbstractBackgroundAtrial fibrillation (AF) is linked to significant morbidity and mortality, with ischaemic stroke being a leading cause of death. Identifying modifiable risk and protective factors may help reduce stroke incidence in patients with AF.MethodsThis umbrella review evaluated meta-analyses of observational studies and randomised controlled trials (RCTs) to assess the association of protective and risk factors with stroke and transient ischaemic attack (TIA) in patients with AF. Observational associations were graded with the Ioannidis framework. Associations from RCTs and non-randomised baseline factors in meta-analyses of RCTs were assessed with GRADE (Grading of Recommendations Assessment, Development and Evaluation). Composite risk scores were summarised separately as risk-stratification tools.Results35 studies were included, comprising 23 meta-analyses of observational studies, reporting on 45 associations and 22 meta-analyses of RCTs on 24 associations based on data from over 7 276 000 participants. Among observational studies, only high levels of N-terminal pro-brain Natriuretic Peptide (NT-proBNP) provided convincing evidence of stroke risk. Previous stroke/TIA and age (65–74 years) were also associated with stroke, supported by a highly suggestive level of evidence. Among meta-analyses of non-randomised factors, female sex, kidney failure, non-paroxysmal AF and type 2 diabetes mellitus were risk factors with moderate to high evidence. However, other well-established risk factors, such as hypertension and vascular disease, were associated with stroke; however, they were supported with a weak level of evidence.ConclusionsDespite stroke being the most severe complication of AF, this umbrella review reveals that few risk factors are supported by high-level evidence. Our findings confirm that elevated NT-proBNP, age and prior stroke are credible stroke risk factors in patients with AF. However, risk factors with weaker evidence, such as hypertension and vascular disease, require further investigation to clarify their actual impact.PROSPERO registration numberCRD42023471263. WHAT IS ALREADY KNOWN ON THIS TOPIC Atrial fibrillation (AF) is associated with significant morbidity, with ischemic stroke being a leading cause of death. Atrial fibrillation (AF) is associated with significant morbidity, with ischemic stroke being a leading cause of death. WHAT THIS STUDY ADDS This is the first study to rank evidence from meta-analyses on stroke and TIA risk factors in AF. Although many factors have been associated with stroke in AF, only a few are well supported: elevated NT-proBNP (strong evidence) and non-paroxysmal AF (moderate evidence). Neither is yet incorporated into current guidelines. In contrast, commonly used risk factors like hypertension and vascular disease show weaker evidence, raising questions about their role in anticoagulation decisions. This is the first study to rank evidence from meta-analyses on stroke and TIA risk factors in AF. Although many factors have been associated with stroke in AF, only a few are well supported: elevated NT-proBNP (strong evidence) and non-paroxysmal AF (moderate evidence). Neither is yet incorporated into current guidelines. In contrast, commonly used risk factors like hypertension and vascular disease show weaker evidence, raising questions about their role in anticoagulation decisions. HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY The findings suggest that anticoagulation guidelines could be improved by prioritizing well-supported risk factors while reconsidering the emphasis on those with limited evidence in current scoring systems. The findings suggest that anticoagulation guidelines could be improved by prioritizing well-supported risk factors while reconsidering the emphasis on those with limited evidence in current scoring systems. Introduction Atrial fibrillation (AF) is the most common sustained arrhythmia worldwide, with high prevalence in adults.1 AF is associated with substantial morbidity and mortality, with stroke being an important cause of death and debilitation.2 The utilisation of oral anticoagulants markedly diminishes the risk of stroke, with a substantial accompanying risk of bleeding.35 Acknowledged risk factors for stroke in AF include heart failure, arterial hypertension, advancing age, diabetes mellitus, a history of transient ischaemic attack (TIA)/stroke and vascular disease. This can be summarised by the CHA2DS2-VA risk score, which, together with its precursor, the CHA2DS2-VASc risk score, is the most used score to guide anticoagulation therapy in patients with AF according to current guidelines.5 However, this score does not consider other potential risk factors, such as biomarkers like N-Terminal Pro-B-type Natriuretic Peptide (NT-pro-BNP),6 rhythm control therapies (antiarrhythmic drugs (AAD)7 and catheter ablation8), AF patterns or AF-burden,9 all of which are linked to stroke risk but not yet used in clinical practice. A thorough evaluation of the evidence concerning risk and protective factors for stroke in AF can guide treatment decisions. However, many underlying studies are fairly dated,10 and the CHA2DS2-VA score has never demonstrated a discriminative ability better than moderate at best.11 Furthermore, there has not been a comprehensive effort to systematically summarise, evaluate and grade the evidence of these risk factors collectively. This study aims to evaluate and synthesise current evidence on risk and protective factors associated with stroke in patients with AF, using the method of the umbrella review.12 Umbrella reviews, previously applied across various clinical fields,13 14 rank the evidence from available meta-analyses on a topic based on sample size, strength of association and potential biases. Methods This umbrella review of meta-analyses adhered to the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) reporting guidelines15 and the MOOSE (Meta-Analysis of Observational Studies in Epidemiology) guidelines.16 The protocol of this study is registered in PROSPERO (CRD42023471263). Data selection, search strategy and selection criteria A comprehensive systematic search was conducted across four electronic databases, PubMed, Web of Science, the Cochrane Database of Systematic Reviews and the Cochrane Central Register of Controlled Trials, from inception through 6 November 2023 to identify meta-analyses of observational studies or randomised controlled trials (RCTs) examining associations between risk or protective factors and the risk of stroke or TIA in patients with AF (Full strategies in online supplemental appendix 1). Additionally, we manually screened the reference lists of all eligible meta-analyses to identify any further relevant studies. Databases were independently searched by two researchers (AK and EC), and eligibility criteria were assessed. The same researchers retrieved articles and further examined them for eligibility. Possible disagreements were settled by consulting another researcher (DT). Peer-reviewed meta-analyses of observational studies with a cohort, case or nested case–control study design or RCTs measuring any association between protective or risk factors and risk of stroke in patients with AF were included. In cases where multiple meta-analyses were available on the same topic, preference was given to the meta-analysis, including a larger number of primary studies.17 18 All relevant outcomes and subgroup analyses from the included original meta-analyses were compiled and reported. Exclusion criteria consisted of (1) employed study designs other than the ones described above (ie, cross-sectional); (2) studies that lacked a systematic review and selection of studies or RCTs; (3) studies that examined genetic risk and protective factors of stroke in patients with AF; (4) studies not published in English; (5) studies providing insufficient data for quantitative analysis; or (6) studies that presented other effects estimates than OR, HR, risk ratio (RR) as, for example, mean difference. Reasons for exclusion after full-text assessment are detailed in online supplemental appendix 2. Data extraction and quality assessment Two researchers (AK and EC) independently extracted data from each eligible meta-analysis. Any discrepancies were resolved through consensus. Variables collected for each study were the name of the first author, the standard identifier (DOI), journal, study design, publication year, number of original studies included, sample size and factors appraised. For each component study, variables such as first author name, publication year, study design, total sample size including exposure and non-exposure population and risk estimates, for example, HR, OR and RR with the 95% CI, were collected. The quality of the methodology in the included meta-analyses was assessed using the AMSTAR2 tool (Assessment of Multiple Systematic Reviews tool, available at https://amstar.ca/Amstar-2.php) by two researchers (AK and EC).19 Risk of bias assessment was performed by two researchers (AK and EC) and checked by two others (DT and DV). Data synthesis and analysis The effect sizes (ES) from each meta-analysis were extracted for every identified association, and pooled effect sizes, along with 95% CIs, were recalculated using random-effects models with the restricted maximum likelihood variance.20 Interstudy heterogeneity was checked with the I2 statistic.21 To evaluate small-study effect bias, the Egger regression asymmetry test was employed.22 The presence of excess significance bias (ESB) was calculated by using the new Test for Excess Statistical Significance and the Proportion of Statistical Significance Test.23 Composite clinical scores (CHA₂DS₂-VASc/CHADS₂/ATRIA) were summarised in separate tables as risk-stratification tools validations. Data analysis was conducted using the ‘metaumbrella: Umbrella Review Package for R’ (V.4.2.2, R Core Team 2023, R Foundation for Statistical Computing, Vienna, Austria; available at https://www.R-project.org).20 A p value of<0.05 was considered statistically significant. In line with previous studies, associations from observational studies were categorised into five levels based on the strength of evidence: convincing (Class I), highly suggestive (Class II), suggestive (Class III), weak (Class IV) and not significant (NS)24 (online supplemental appendix 3, eTable 1). We applied the same certainty grading (high/moderate/low/very low) to both randomised interventional meta-analyses and analyses of non-randomised baseline factors within trial cohorts. Judgements considered risk of bias, inconsistency, indirectness, imprecision and publication bias using the Grading of Recommendations Assessment, Development and Evaluation (GRADE) algorithm implemented in the metaumbrella package.20 25 This preserves a consistent certainty scale while avoiding the impression of RCT-level evidence for non-randomised exposures. Every analysis begins at a ‘High’ rating and is downgraded by one or two levels when predefined criteria indicate publication bias (Egger test), inconsistency (I²), imprecision (low statistical power) or study limitations (risk of bias). The quantitative thresholds followed, together with the full numerical outputs for each analysis, are provided in online supplemental appendix 3, eTable 2, and the accompanying explanatory text.20 This algorithmic approach differs from the methodology in the standard GRADE framework.26 We assessed overlap among meta-analyses using Corrected Covered Area (CCA) and Jaccard % metrics27 (online supplemental appendix 6). Results Literature search First, 1491 publications were identified in MEDLINE/PubMed database. After removing duplicates and screening titles and abstracts, 93 articles remained. Subsequently, following full-text evaluation, 58 articles were excluded (online supplemental appendix 2). Ultimately, 35 meta-analyses were included in the final analysis. Among these, 23 assessed observational studies and included 45 associations, while 22 RCTs were assessed, encompassing 24 associations (figure 1; tables13; online supplemental eFigure 1 and 2, appendix 4). We summarised the meta-analyses of risk scores in separate tables. As expected for tools that aggregate multiple predictors, higher thresholds were associated with higher stroke risk; these findings reflect predictive performance rather than causal inference for any single component. Flowchart of the study selection process. Table 1Risk/protective factors, with convincing, highly suggestive and suggestive evidence (Class I, II, III) in meta-analyses of observational studiesAuthor, yearRisk/protective factorKn/NMetricES (95% CI)P valuePI include null valueI2SSEESBLS signCEAMSTAR 2 qualityAhmadzadeh et al6High NT-ProBNP1112 460/11 939HR1.825(1.458 to 2.285)1.48E−07No10.6%NoNoNoILowNoubiap, 2021Previous stroke/TIA1568 555/1 484 851HR2.836(2.189 to 3.674)3.01E−15Yes95.3%NoNoYesIIModerateNoubiap, 2021Age 65–74114511/98 815HR2.143(1.856 to 2.474)2.58E−25No75.6%NoNoYesIIModerateSagris et al9High BNP/NT-proBNP65987/75 848RR1.917(1.494 to 2.458)3.02E−07Yes52.8%NoNoYesIILowNoubiap, 2021Age ≥75158569/121 515HR2.826(2.276 to 3.507)4.46E−21No89.9%YesNoNoIIIModerateNoubiap, 2021Heart failure1725 393/449 606HR1.248(1.109 to 1.404)0.000238Yes88.3%NoYesNoIIIModerateWagstaff, 2014/Noubiap, 2021Female2818 280/345 472OR1.224 (1.088 to 1.377)000078Yes83.1%NoYesNoIIICritically low/moderateLiu, 2020High BMI94050/82 631RR0.908(0.859 to 0.96)0.000688Yes39.5%NoNoNoIIILowBuhari et al38Early rhythm control340 992/320 596RR0.748(0.659 to 0.849)0.00000678Yes77.7%NoNoYesIIICritically lowLilli, 2017Paroxysmal AF47217/111 768RR0.729(0.64 to 0.83)0.00000168No0%NoNoYesIIIModerateAF, atrial fibrillation; AMSTAR, Assessment of Multiple Systematic Reviews; BMI, body mass index; BNP, brain natriuretic peptide; CE, class of evidence; ES, effect size; ESB, excess significance bias; I2, heterogeneity; K, number of studies for each factor; LS, largest study with significant effect; n, number of cases; N, number of controls per factor; NT-proBNP, N-terminal pro-brain Natriuretic Peptide; PI, prediction interval; RR, risk ratio; SSE, small study effects; TIA, transient ischemic attack. Risk/protective factors, with convincing, highly suggestive and suggestive evidence (Class I, II, III) in meta-analyses of observational studies AF, atrial fibrillation; AMSTAR, Assessment of Multiple Systematic Reviews; BMI, body mass index; BNP, brain natriuretic peptide; CE, class of evidence; ES, effect size; ESB, excess significance bias; I2, heterogeneity; K, number of studies for each factor; LS, largest study with significant effect; n, number of cases; N, number of controls per factor; NT-proBNP, N-terminal pro-brain Natriuretic Peptide; PI, prediction interval; RR, risk ratio; SSE, small study effects; TIA, transient ischemic attack. Table 2Risk/protective factors, with weak evidence (Class IV) in meta-analyses of observational studiesAuthor, yearRisk/protective factorKn/NMetricES (95% CI)P valuePI include null valueI2SSEESBLS signCEAMSTAR 2 qualityNoubiap, 2021Age ≥853620/8477HR6.868 (6.334 to 7.446)0.00000No0%NoNAYesIVModerateSagris, 2021AHRE ≥30 s341/663RR6.002(3.01 to 11.967)3.59E–07Yes0%NoNoYesIVModerateSagris, 2021AHRE ≥6 hours135/2420RR3.14(1.49 to 6.62)0.00263NANANANAYesIVModerateSagris, 2021AHRE ≥24 hours3703/24 245RR2.532 (1.315 to 4.873)0.00542Yes61.3%NoNoYesIVModerateSagris, 2021AHRE ≥5 mins8813/24 492RR2.067(1.456 to 2.934)0.0000484Yes38.6%NoYesYesIVModerateCorica, 2022Cerebral microbleeds6169/3630OR1.774(1.262 to 2.496)0.000982No16.2%NoNoNoIVModerateYang, 2022AF burden>5 min11963/36 603RR1.673(1.246 to 2.247)0.000627Yes52.0%NoYesYesIVLowZeng, 2015eGFR<60122885/55 982RR1.65 (1.194 to 2.279)0.00238Yes77.4%NoYesNoIVCritically lowZhu, 2016Underweight36134/884 985RR1.638(1.055 to 2.543)0.0281Yes18.2%NoNoNoIVHighNoubiap, 2021Hypertension16876/19 069HR1.597(1.371 to 1.86)1.83E–09Yes87.7%YesYesYesIVModerateGanesan, 2016NPAF2870/26 636OR1.411(1.006 to 1.978)0.046NA0%NANANoIVLowNoubiap, 2021Diabetes mellitus18494/15 509HR1.281(1.198 to 1.371)6.21E–13No62.1%YesYesNoIVModerateTeppo, 2021Mental health condition748 602/798 676HR1.252(1.082 to 1.45)0.0026No0%NoNoNoIVCritically lowNoubiap, 2021Vascular disease15966/30 371HR1.214(1.057 to 1.393)0.00607Yes70.9%YesYesYesIVModerateSachdeva, 2021High atrial strain6537/5072OR0.876(0.801 to 0.957)0.00347Yes85.5%NoYesYesIVCritically lowZhu, 2016Obese81395/12 292RR0.84(0.722 to 0.977)0.0233Yes0%NoNoNoIVHighAF, atrial fibrillation; AHRE, atrial high rate episodes; AMSTAR, Assessment of Multiple Systematic Reviews; CE, class of evidence; eGFR, estimated glomerular filtration rate; ES, effect size; ESB, excess significance bias; I2, heterogeneity; K, number of studies for each factor; LS, largest study with significant effect; N, number of controls per factor; n, number of cases; NA, not assessable; NPAF, non-paroxysmal atrial fibrillation; PI, prediction interval; RR, risk ratio; SSE, small study effects. Risk/protective factors, with weak evidence (Class IV) in meta-analyses of observational studies AF, atrial fibrillation; AHRE, atrial high rate episodes; AMSTAR, Assessment of Multiple Systematic Reviews; CE, class of evidence; eGFR, estimated glomerular filtration rate; ES, effect size; ESB, excess significance bias; I2, heterogeneity; K, number of studies for each factor; LS, largest study with significant effect; N, number of controls per factor; n, number of cases; NA, not assessable; NPAF, non-paroxysmal atrial fibrillation; PI, prediction interval; RR, risk ratio; SSE, small study effects. Table 3Risk/protective factors of non-randomised baseline factors in meta-analyses of randomised control trials with significant associationsAuthor, yearRisk/protective factorKn/NMetricES (95% CI)P valueI2SSERoBGLEAMSTAR 2 qualityPancholy, 2014Female_VKA5349/9849OR1.279(1.111 to 1.473)6.08e–040%No100%HighCritically lowNoubiap, 2021Female175087/167 924OR1.208(1.13 to 1.292)3.40e–089%No100%HighModerateZeng, 2015eGFR<6051666/61 786RR1.797(1.516 to 2.13)1.42e–1151.7%No100%ModerateCritically lowAhmadzadeh, 2023High NT-ProBNP871 948/142 029HR1.795(1.513 to 2.129)2.02e–1137.1%No100%ModerateCritically lowGanesan, 2016NPAF119641/272 976OR1.357(1.217 to 1.513)3.74e–0871.1%No100%ModerateLowAbdool, 2020Diabetes mellitus31094/32 377RR1.184(1.016 to 1.38)3.04e–0259.5%No100%ModerateCritically lowMahajan, 2018SubClinical AF261/2892OR5.658(4.015 to 7.973)4.13e–230%NA100%WeaklowLiu, 2020High BMI1381/18 113RR0.778(0.702 to 0.862)1.53e–06NANA100%WeakLowSagris, 2021AHRE ≥5 mins4128/3768RR2.098(1.428 to 3.082)1.61e–040%No0%Very weakModerateSagris, 2021AHRE ≥10 s1119/–2718RR0.88(0.55 to 1.41)5.96e–01 (NS)NANA0%Very weakModerateIsxyeiuiiChen, 2015/Lilli, 2017Paroxysmal AF205995/199 358RR0.715(0.647 to 0.789)3.33e–1162.0%No6.61%Very weakCritically low/moderateNon-randomised baseline factors analysed within trial cohorts are synthesised as trial-cohort prognostic meta-analyses and graded with GRADE-Prognosis are presented here. Effects were pooled using random-effects models; I² quantifies between-study heterogeneity. Certainty reflects risk of bias, inconsistency, indirectness, imprecision and publication bias.AF, atrial fibrillation; AHRE, atrial high rate episodes; AMSTAR, Assessment of Multiple Systematic Reviews; BMI, body mass index; eGFR, estimated glomerular filtration rate; ES, effect size; GLE, GRADE level of evidence; GRADE, GRADE, Grading of Recommendations Assessment, Development and Evaluation; I2, heterogeneity; K, number of studies for each factor; N, total number of cohort per factor; n, number of cases; NA, not assessable; NPAF, non-paroxysmal AF; NT-proBNP, N-terminal pro-brain Natriuretic Peptide; PI, prediction interval; RR, risk ratio; SSE, small study effects; VKA, vitamin K antagonist. Risk/protective factors of non-randomised baseline factors in meta-analyses of randomised control trials with significant associations Non-randomised baseline factors analysed within trial cohorts are synthesised as trial-cohort prognostic meta-analyses and graded with GRADE-Prognosis are presented here. Effects were pooled using random-effects models; I² quantifies between-study heterogeneity. Certainty reflects risk of bias, inconsistency, indirectness, imprecision and publication bias. AF, atrial fibrillation; AHRE, atrial high rate episodes; AMSTAR, Assessment of Multiple Systematic Reviews; BMI, body mass index; eGFR, estimated glomerular filtration rate; ES, effect size; GLE, GRADE level of evidence; GRADE, GRADE, Grading of Recommendations Assessment, Development and Evaluation; I2, heterogeneity; K, number of studies for each factor; N, total number of cohort per factor; n, number of cases; NA, not assessable; NPAF, non-paroxysmal AF; NT-proBNP, N-terminal pro-brain Natriuretic Peptide; PI, prediction interval; RR, risk ratio; SSE, small study effects; VKA, vitamin K antagonist. Overall, CCA was slight in both streams (observational 1.18%; RCT 0.50%). However, a few factor clusters showed>10% overlap (online supplemental appendix 6). Meta-analyses of observational studies The median number of studies comprised in included meta-analyses was 6 (IQR=4–11), with a median sample size of 24 407.5 (IQR=16 002–118 985), and a median of 2885 events (IQR=659–8569). According to AMSTAR2, the methodological quality of the included meta-analyses of observational studies was rated as high in one case, moderate in five, low in nine and critically low in eight. Among the meta-analyses of observational studies, 32 of the 45 studied associations (71%) showed a nominally statistically significant effect (p≤0.05) under random-effects models, with 11 (24%) achieving a p value<10−6. 23 associations (51%) exhibited substantial heterogeneity (I2>50%), and only 10 associations (22%) had a 95% prediction interval that excluded the null value. In 19 associations (42%), the ES of the largest study showed a nominally statistically significant effect (p≤0.05). Finally, small study effects were detected for 7 associations (16%), and ESB was identified for 10 (22%). Applying the credibility of evidence criteria, only one (2%) association presented convincing evidence (Class I); high NT-proBNP values. Four additional associations (8.9%) showed highly suggestive evidence (Class II) for certain risk factors: an age range between 65 and 74 years, history of previous stroke or TIA, elevated values of either BNP or NT-proBNP and a low-risk Anticoagulation and Risk Factors in Atrial Fibrillation (ATRIA) score28 (compared with low-risk CHA2DS2-VASc score). Five associations were suggestive of risk factors such as older age≥75 years, CHA2DS2-VASc≥2, high-risk ATRIA score (compared with high-risk CHA2DS2-VASc score), female sex and heart failure. Lastly, other associations presented suggestive evidence for protective factors, including early rhythm control (anti-arrhythmic medication or catheter ablation started within 12 months of AF diagnosis as compared with rate-control therapy), high body mass index and paroxysmal AF compared with non-paroxysmal AF (NPAF) (table 1, figure 2). The remaining 19 (42%) statistically significant associations between risk or protective factors and risk of stroke presented weak evidence, while 13 associations (29%) had no evidence (tables2 3, figure 2; Online supplemental appendix 5, eTable 1 and 4). CHA2DS2-VASc score and ATRIA scores reflect tool performance, not an independent risk factor. Forest plot of risk/protective factors in meta-analyses of observational studies. AF, atrial fibrillation; AHRE, atrial high rate episodes; BMI, body mass index; BNP, brain natriuretic peptide; eGFR, estimated glomerular filtration rate; I2, heterogeneity; NPAF, non-paroxysmal AF; NT-proBNP, N-terminal pro-brain Natriuretic Peptide; TIA, transient ischaemic attack. Analyses of non-randomised baseline factors within trial cohorts The methodological quality of the trial-cohort prognostic meta-analyses by AMSTAR2 was moderate in three, low in six and critically low in nine meta-analyses (table 3, figure 3, online supplemental appendix 5). Across associations, the median number of studies was 3 (IQR 2–6.5), the median sample size was 25 804 (IQR 6809–101 908), and the median number of cases was 848 (IQR 185–3377). Overall, 11/17 (65%) associations were nominally significant at p<0.05, 9/17 (53%) at p<0.001. Substantial heterogeneity (I²>50%) was present in 6/15 (40%) associations with I² reported; small-study effects were not detected in this subset. For certainty, using GRADE, two associations were rated high (female sex overall; female on vitamin K antagonist (VKA)), four were moderate (high NT-proBNP, diabetes mellitus, estimated glomerular filtration rate<60, NPAF) and the remaining significant signals were weak/very weak; the rest were not statistically significant. The quality of included RCT meta-analyses according to AMSTAR2 was rated as moderate for 4 meta-analyses, low for 6 and critically low for 12 (table 3, figure 3 and online supplemental appendix 5, eTable 2 and 5). Forest plot of risk/protective factors in meta-analyses of non-randomised baseline factors within trial cohorts and of randomised controlled trials. Note: Entries based on randomised factors in RCT meta-analyses are indicated by (RCT) in brackets. AF, atrial fibrillation; AHRE, atrial high rate episodes; BMI, body mass index; eGFR, estimated glomerular filtration rate; I2, heterogeneity; NPAF, non-paroxysmal AF; NT-proBNP, N-terminal pro-brain Natriuretic Peptide; RCT, randomised controlled study; VKA, vitamin K antagonist. Meta-analyses of randomised controlled studies When restricting RCT meta-analyses to those in which the exposure was randomised, only two associations were available: pharmacologic rate versus rhythm-control and catheter ablation versus non-ablation/medical therapy. The first association was not significant, while the latter showed weak GLE (online supplemental appendix 5, eTable 3). Discussion To our knowledge, this is the first umbrella review to examine risk and protective factors for stroke/TIA in patients with AF, across 69 risk and protective factors, identifying 44 significant associations in over 7 276 000 participants. Only high NT-proBNP levels showed a significant association with stroke, by convincing evidence in meta-analyses of observational studies and by moderate evidence in non-randomised factors in RCT meta-analyses. Notably, well-established risk factors such as hypertension, vascular disease and heart failure, which are included in the CHA2DS2-VASc score, displayed weak evidence for their associations with stroke. Similarly, a CHA2DS2-VASc=1 also showed weak correlation; however, this reflects tool performance and not an individual risk factor. In contrast, factors, such as NT-proBNP (convincing evidence) and NPAF (moderate evidence), are not considered in current risk scores. High NT-proBNP levels were the strongest risk predictor for stroke, OR 1.82 (95% CI 1.46 to 2.29). NT-proBNP is associated with the risk of stroke, independently of other risk factors and conditions29 including ongoing AF during which NT-proBNP is usually elevated.30 NT-proBNP measurements can, therefore, improve risk stratification beyond the CHA2DS2-VA score and might be a novel tool for improved stroke prediction in AF.31 Among the strongest predictors of ischaemic stroke in AF in our study is advanced age. Our analysis revealed that applying higher age cut-offs significantly increases the OR, ranging from 2.14 (95% CI 1.86 to 2.47) for individuals aged 65–74 to 6.87 (95% CI 6.33 to 7.45) for those over 85. There is evidence suggesting that age≥55 years might also be a risk factor for stroke among Asian patients with AF.32 However, despite strong supportive evidence, older people with AF are less likely to receive oral anticoagulation (OAC) therapy. In addition, older patients (≥75 years) were under-represented in OAC versus placebo RCTs. Even though age is a non-modifiable risk factor,33 34 it is an important consideration when making decisions about anticoagulation.35 The type of AF and its association with the risk of stroke has been controversial. A report from a Swedish cohort has shown that AF patients with paroxysmal and NPAF had similar incidences of stroke.36 However, numerous studies, including big RCTs such as ARISTOTLE33 and ROCKET-AF,34 showed that patients with NPAF had a higher risk of stroke compared with those with paroxysmal AF. These data agree with the results shown in this study; NPAF elevates the risk for stroke with a moderate level of evidence from meta-analyses, including RCT data. Female patients with AF had a higher stroke risk, supported by high GLE in RCT meta-analyses and suggestive evidence in observational syntheses. The VKA-specific association is context-limited to VKA therapy and not generalised to direct-acting oral anticoagulants (DOAC)-treated populations. Further, recent studies suggest that female sex is a risk modifier for AF-associated stroke rather than an independent risk factor.37 Consequently, it has been removed from the CHA2DS2-VASc score in the recent European Society of Cardiology guidelines for AF.5 The differences in stroke risk between sexes can be attributed to older age and discrepancies in cardiovascular care,38 39 highlighting the importance of comprehensive risk factor management, especially in older women with AF. Notably, nationwide registry data and pooled DOAC trial analyses indicate that the excess stroke risk in women disappears when anticoagulation is adequate, either with DOACs or VKA therapy maintained at time-in-therapeutic-range ≥ 65%.37 40 41 This supports the view that female sex mainly reflects modifiable factors such as undertreatment and comorbidity rather than an intrinsic biological risk. Another interesting factor showing an increased risk of stroke with medium GLE in RCTs is renal impairment in patients with AF. Renal impairment in patients with AF has been linked to a higher risk of stroke, comparable to established factors like heart failure, hypertension, advanced age and diabetes. Chronic kidney disease likely aggravates stroke risk in AF by intensifying prothrombotic conditions through various physiological mechanisms.42 It is worth noting that several commonly acknowledged risk factors included in the CHA2DS2-VA score, such as hypertension and vascular disease, had weak supporting evidence for these associations. An explanation for these findings could be that they are based on older studies or because stroke and comorbidities like hypertension can coexist.43 Some studies using the Mendelian randomisation design suggest a potential causal relationship between factors like hypertension and stroke.44 However, based on our results, more data are needed to provide more robust evidence for individual CHA2DS2-VA risk factors generating one point, apart from age, and their association with stroke. The need to identify the most significant risk and protective factors for stroke in patients with AF is crucial for proper treatment. Current guidelines5 recommend using the CHA2DS2-VA score. This risk score and its precursor (CHA2DS2-VASc) have been important in the increase in stroke preventive treatment in patients with AF. However, there is always room for improvement. While some components of the CHA₂DS₂-VA score (eg, age≥65, prior stroke/TIA) were supported by strong evidence, others (eg, hypertension, vascular disease) had weaker meta-analytical support. Incorporating additional biomarkers (eg, NT-proBNP) or AF type (NPAF) may improve stroke risk stratification. This umbrella review of meta-analyses, while comprehensive, has several limitations. First, our inclusion depended on existing meta-analyses at the time of the search cut-off. Several plausible AF-stroke determinants remain without AF-restricted meta-analyses, particularly cumulative AF burden, atrial structural/functional markers and inflammatory/prothrombotic biomarkers, summarised in an evidence gap map (online supplemental appendix 7). Second, our quality grading reflects the design of the meta-analysis (observational, RCT or non-randomised factor in RCT cohort) rather than the randomised or non-randomised nature of individual associations. Associations for non-randomised factors, although estimated within RCT datasets, were graded with GRADE as observational evidence because they arise from secondary analyses where the exposure was not randomised. They should be interpreted accordingly. Third, the results from the primary analyses may have been influenced by various unidentified covariates, potentially leading to skewed conclusions. Fourth, potentially meaningful subgroup analyses for different age groups, genders, use of anticoagulants and types of AF were addressed separately only when individual meta-analyses provided sufficient data. Fifth, the included meta-analyses varied in quality. While most associations graded ‘convincing’ or ‘high’ came from reviews that achieved adequate AMSTAR2 ratings, findings drawn from lower-rated reviews are less reliable and should be interpreted with caution. This underscores the need to re-evaluate the evidence base as higher-quality systematic reviews become available. Sixth, despite conducting Egger’s test for small-study effects, it remains a risk for publication bias as studies with non-significant or negative results may not have been published or captured in our search. Seventh, the grading system can only indicate systematic biases, but no proof of their nature and extent. Finally, our analysis assessed individual risk factors, the CHA₂DS₂-VA score’s discriminative performance and clinical validity were beyond the scope of this umbrella review. Although overall CCA was slight, we observed non-trivial duplication in a few topic clusters (pairs>10% overlap), for which precision should be interpreted cautiously. Conclusions Stroke is the most dangerous complication of AF. This study reveals that only a few associations between protective and risk factors with stroke/TIA are supported by high-level evidence. High NT-proBNP has convincing observational evidence (and moderate in trial-cohort analyses), age 65–74 and prior stroke/TIA are highly suggestive, NPAF shows moderate evidence. Early rhythm control and paroxysmal AF show suggestive protective associations. Several CHA₂DS₂-VA components (eg, hypertension, vascular disease) have weak support and warrant re-evaluation. Incorporating NT-proBNP and AF phenotype may improve risk stratification beyond current scores. Supplementary material