Our study aimed to examine the prevalence of ESUS and analyze factors associated with stroke severity and recurrence. The real-world study of the embolic stroke of undetermined sources (ROS-ESUS) cohort was named based on its construction and comprised patients who were admitted to each stroke center between January 2014 and December 2019 with ischemic strokes classified as ESUS. This nationwide study, encompassing 20 centers, received approval from the institutional review boards of all participating institutions in 2022. ROS-ESUS cohort enrollment was considered based on the presence of non-lacunar infarction occurring within 7 days in patients aged ≥20 years with undetermined etiology, under the criteria of Trial of Org 10172 in Acute Stroke Treatment (TOAST) classification as follows: negative evaluation (or cryptogenic stroke) or cardioembolism, but not high-risk cardioembolic sources [9,10]. The overall ESUS classification was determined by physicians at each stroke center, and all patients classified as having ESUS were consecutively enrolled in the ROS-ESUS cohort.

The exclusion criteria included: (1) patients diagnosed with atrial fibrillation, (2) those with active cancer, (3) those with high risk of cardioembolism, (4) those with last normal time of 7 days or more, and (5) those who the attending physician determined that the patient is not an appropriate candidate based on clinical judgment. Detailed clinical imaging (transthoracic echocardiography) and laboratory data were collected from each center. The inclusion flowchart is shown in Figure 1.

Patients who met the eligibility criteria were followed up at 3 months, 1 year, and beyond 1 year after their diagnosis of ESUS. Follow-up was discontinued in cases of loss to follow-up or death. Although the maximum follow-up duration varied slightly across institutions, most patients were followed up until 2023, when possible.

The study design was approved by the Institutional Review Board of Samsung Medical Center (IRB No. SMC-2022-02-010). Due to its retrospective nature, the requirement for written informed consent was waived.

AC was defined based on the presence of either of the following criteria: (1) atrial enlargement was defined as having a left atrial (LA) diameter exceeding 40 mm for men and 38 mm for women on echocardiography (parasternal long axis view) performed during the time of admission by each institute [11]. In addition, participants with LA volume index exceeding 34 mL/m² were classified as having atrial enlargement [12], and (2) elevated N-terminal pro-B-type natriuretic peptide (NT-proBNP) levels of ≥250 pg/mL during admission were considered indicative of atrial dysfunction [13]. We also classified each individual according to the number of criteria they met. They were grouped not only by the presence of AC (AC and non-AC groups) but also by the number of factors that were satisfied (groups 0, 1, and 2). Patients without NT-proBNP measurements (n=3,845) and echocardiographic features (n=695) were classified as not meeting the diagnostic criteria.

The primary outcome of this study was the time-to-first recurrent ischemic stroke. Patients with hemorrhagic strokes, transient ischemic attacks (TIAs), or strokes of unidentifiable type were excluded. The time-to-major major adverse cardiovascular events (MACEs) was the secondary outcome. MACE comprised recurrent ischemic stroke, hemorrhagic stroke, symptomatic heart failure, myocardial infarction, and vascular death.

Ischemic stroke severity was assessed using the National Institutes of Health Stroke Scale (NIHSS) scores, changes in modified Rankin Scale (mRS) scores, and the length of hospital stay was also analyzed.

Statistical analyses were conducted using R (version 4.4.1; R Foundation for Statistical Computing, Vienna, Austria) and Python (version 3.12.4; https://www.python.org/), with a significance threshold set at P<0.05. Baseline characteristics between groups, categorized by the presence of AC, were compared using the chi-square test and the Mann-Whitney U test. Kaplan-Meier survival analysis, along with log-rank tests, was performed to assess the survival probability between AC groups. Competing risk analysis was then employed with the Fine-Gray model to compare outcomes of recurrent strokes and MACEs while accounting for unrelated deaths. The covariates considered were age, sex, and body mass index (BMI), along with medical history of stroke, TIAs, peripheral arterial disease, coronary heart disease, hypertension, diabetes mellitus, dyslipidemia, and smoking status, which were selected as prominent risk factors for stroke [14]. They were then independently subjected to regression analysis to assess vascular outcomes, and those meeting the P<0.10 were primarily considered for each multivariate analysis. Although most collected data were not modified, BMI measurements out of three standard deviation ranges were substituted with the median values. Age and BMI data were standard-scaled for statistical analyses, and missing data for covariates were rare and were therefore excluded from the statistical analysis.

Stroke severity was assessed using NIHSS scores at admission (baseline) and discharge. For analysis, severity levels were classified into three groups: mild (0-4), moderate (5-14), and severe (≥15), and comparisons were conducted using chi-square tests and residual heatmaps. Ordinal variables, including NIHSS scores and mRS scores, were analyzed by comparing medians and interquartile range (IQR), while continuous variable, length of hospitalization, was analyzed by comparing means and standard deviations. Mann-Whitney U or Kruskal-Wallis H tests were applied, followed by Dunn’s test with Bonferroni correction when necessary. Two patients without admission NIHSS scores were excluded from the analysis, and hospitalized days outside of the three standard deviation ranges were substituted with the median value.

A propensity score (PS)-matched dataset was created to ensure balance between the AC and non-AC groups. The PS was derived using a multivariate logistic regression model including age, sex, BMI, admission NIHSS score, CHA₂DS₂-VASc score, and medical history variables such as congestive heart failure, hypertension, diabetes mellitus, history of TIA and stroke, peripheral arterial disease, coronary heart disease, dyslipidemia, and smoking, as independent predictors. Blood pressure, pulse rate, presence of atrial fibrillation during admission, and use of antiplatelet agents or oral anticoagulants at discharge were also recorded. The variables were carefully selected to maximize randomization after reviewing factors associated with stroke recurrence [15]. Missing values (less than 1%) were replaced with the median value of each variable. PSM was performed using the nearest-neighbor method with 1:1 matching within a caliper of 0.12. The performance of PS matching was assessed by quantifying the absolute standardized differences in patient characteristics (Supplementary Figure 1). After, univariate Cox regression was performed to compare the AC and non-AC groups, as well as the stratified groups. Finally, sensitivity analysis was employed to determine the trends in recurrent ischemic stroke. All graphs exhibited linear trends in both directions or parabolic trends. Although the significance was mixed, as the presence of AC alone shows marginal significance, the result was consequential to the stratified group sensitivity with strong significance. Therefore, the tendency was numerically comparable.

Among the 5,787 patients investigated, the mean age and BMI were 65.9±13.9 years (female 39.8%) and 23.7±3.2, respectively. While 2.1% of patients had a history of TIA, 19.0% had a history of stroke, including hemorrhagic stroke. Comorbidities included hypertension (61.1%), diabetes mellitus (29.5%), dyslipidemia (39.6%), coronary heart disease (10.8%), peripheral arterial disease (1.3%), and current smoking status (26.7%).

The ratios for each condition varied, as 8.7% of the patients had NT-proBNP level ≥250 pg/mL, and 41.1% had LA enlargement based on an increased LA diameter or volume index. A total of 45.0% of patients satisfied our definition of AC, and groups 0, 1, and 2 accounted for 55.0%, 40.3%, and 4.7%, respectively. NIHSS scores at admission and discharge were 3 (1-6) and 2 (0-4), respectively. The average follow-up duration was 16.1±19.1 months, with a maximum of 107.9 months. During this period, 246 patients experienced a recurrent ischemic stroke, and 306 patients had a MACE event (Supplementary Table 1). Comparison between the AC and non-AC groups is presented in Table 1.

The Kaplan-Meier curves showed a marginal difference between the AC and non-AC groups for recurrent stroke (P=0.05). When stratified, group 2 was responsible for greater increase in hazard ratio (HR) (P=0.04) compared to the other groups (Supplementary Figure 2A and B). When unrelated deaths were considered, the AC group exhibited borderline significance in ischemic stroke recurrence in univariate (HR: 1.26, 95% confidence interval [CI]: 0.98-1.62, P=0.07) and multivariate (HR: 1.25, 95% CI: 0.97-1.60, P=0.08) assays. Moreover, while group 1 exhibited low significance in univariate (HR: 1.20, 95% CI: 0.93-1.56, P=0.16) and multivariate (HR: 1.19, 95% CI: 0.92-1.55, P=0.18) analyses, group 2 displayed compelling results in both univariate (HR 1.75, 95% CI: 1.06-2.89, P=0.03) and multivariate (HR: 1.76, 95% CI: 1.06-2.92, P=0.03) models (Table 2).

However, the MACE outcomes were contradictory. The Kaplan-Meier curves also exhibited marginal significance in the AC and non-AC groups (P=0.05), whereas group 2 displayed a larger difference (P=0.03) (Supplementary Figure 2C and D). However, in the subdistribution hazard analysis, all the results were not significant (all P>0.05), even in the stratified group 2 (univariate HR: 1.56, 95% CI: 0.96-2.52, P=0.07) (Table 2).

After PS matching, 2,604 pairs were selected, exhibiting a notable increase in the risk of ischemic stroke in the AC group (P=0.02) and stratified AC group 2 (P=0.01) (Table 2). The difference between curves increased in the original cohort. In MACE, an early phase overlap existed in the AC group versus the non-AC group (P=0.11), but stratified AC group 2 was similar to the ischemic stroke results (P=0.02) (Figure 2).

HR for the AC group was estimated and exhibited significance in ischemic stroke recurrence (HR: 1.37, 95% CI: 1.04-1.79, P=0.02), but not with MACE (HR: 1.21, 95% CI: 0.96-1.53, P=0.11), which was consistent with competing risk analysis models. The AC group was evaluated after stratification into groups 1 and 2. When measured in the PS matched cohort, the significance in ischemic stroke recurrence escalated in groups 1 (HR: 1.30, 95% CI: 0.99-1.73, P=0.06) and 2 (HR: 1.94, 95% CI: 1.16-3.26, P=0.01), although group 1 exhibited borderline significance. MACE outcomes were statistically meaningful in group 2 (HR 1.70, 95% CI: 1.07-2.70, P=0.02), unlike the previous results in the original cohort.

There was no significant difference in stroke severity (mild, moderate, or severe) at baseline or discharge between the AC and non-AC groups (P>0.05) (Figure 3). However, elevated NT-proBNP levels (Spearman’s rank correlation: ρ=0.12, P<0.05) and LA enlargement (median NIHSS score: 3 in patients without LA enlargement vs. 2 in those with enlargement; P<0.05 by Mann-Whitney U test) were both significantly associated with NIHSS. In addition, when the groups were stratified according to the number of AC factors, the effects of the non-AC, AC, and Group 1 groups were mixed, and group 2 displayed a clear correlation with stroke severity (Supplementary Figure 3). No difference in NIHSS score existed between non-AC and AC groups. However, when the AC group was stratified by the number of factors, group 2 surpassed the non-AC group and group 1 in terms of severity during admission (difference in median was both +2, P<0.05) and discharge (difference in median was both +1, P<0.05), whereas group 1 exhibited no difference compared to the non-AC group. Notably, hospitalization stays differed in trends. The AC group surpassed the non-AC group in hospital stays by 0.38 days (P<0.05), and group 1 patients exhibited longer hospital stays than those in the non-AC group by 0.29 days (P<0.05). Group 2 showed an increase in hospitalization by 0.83 (P<0.05) and 1.12 days (P<0.05) compared to group 1 and non-AC group, respectively (Figure 3).

Functional outcomes also supported our findings. In the entire cohort (n=5,787), functional decline—expressed as mRS change (discharge mRS minus admission mRS)—increased in a step-wise fashion: the median mRS change was +1 in non-AC group and group 1, and +2 in group 2 (P<0.05). This gradient remained after restricting the analysis to patients with baseline mRS ≤2 (sensitivity cohort, n=4,913), where median mRS change values were again +1, +1, and +2 across non-AC group, group 1, and group 2, respectively (P<0.05). These findings show that greater AC burden independently predicts larger in-hospital deterioration in functional status.

The contributors of AC (NT-proBNP levels and the presence of LA enlargement) were assessed for their impact on the original cohort. Effects were measured for ischemic stroke recurrence using univariate Cox regression analysis, which did not exhibit significant results in both NT-proBNP levels (HR: 0.97, 95% CI: 0.84-1.12, P=0.69) and LA enlargement (HR: 1.15, 95% CI: 0.89-1.49, P=0.28) individually.

The concept of AC has emerged as a potential mechanism of stroke in patients with ESUS, extending beyond the traditional markers of atrial fibrillation. This shift was supported by several key observations. Structural and functional changes in the atrial myocardium can occur before the development of detectable atrial fibrillation, creating a prothrombotic environment [16,17]. Second, markers of atrial dysfunction such as elevated NT-proBNP levels and LA enlargement have been independently associated with stroke risk, even without documented atrial fibrillation [11,18]. Although our study did not include electrocardiographic features, such as P-wave terminal force in V1 or ectopic atrial activities, they could be considered candidates for diagnosing AC. Nonetheless, our findings align with and extend previous research [6,19], demonstrating that the markers of atrial dysfunction collectively contribute to stroke recurrence in patients with ESUS. In contrast, MACE includes various outcomes, such as ischemic stroke, hemorrhagic stroke, and cardiac disorders, which may explain why AC was not a strong indicator of MACE in our study. However, after PS matching, group 2, despite both factors, exhibited significance, suggesting that the combination of factors contributed to the diverse vascular outcomes.

Our study revealed that approximately 45% of patients with ESUS exhibited signs of AC, while the NAVIGATE-ESUS (Rivaroxaban Versus Aspirin in Secondary Prevention of Stroke and Prevention of Systemic Embolism in Patients With Recent Embolic Stroke of Undetermined Source) and ARCADIA (AtRial Cardiopathy and Antithrombotic Drugs In Prevention After Cryptogenic Stroke) trials reported 37% and 41%, respectively[3,5]. Variations may emerge based on the criteria used to define atrial fibrillation. In our study, the presence of AC was associated with a 1.2-1.9 fold increase in risk of ischemic stroke recurrence, independent of other established risk factors. This association remained robust even after adjusting for potential confounders.

Previous studies have explored the relationship between various components of atrial cardiomyopathy and the recurrence of ischemic stroke. Several studies have linked LA enlargement to stroke recurrence [11,20,21], while others have reported an increased recurrence rate associated with elevated NT-proBNP levels [22,23]. Additionally, other contributing factors, such as the P-wave terminal force in lead V1 [21,23] and the lower LA emptying fraction [24] have been actively investigated. Although the findings of these studies are not always consistent, a general trend indicating that these factors increase the risk of stroke recurrence has been observed. Unlike previous studies, our study demonstrated the potential utility of factor stratification in identifying patients at higher risk in a large sample. Our findings revealed a significant increase in ischemic stroke recurrence when both factors were present, compared to AC alone. This observation suggests that the inconsistent results in previous studies may stem from varied approaches to atrial cardiomyopathy assessment, underscoring the necessity of a comprehensive evaluation of the contributing factors.

The observed increase in recurrence may be explained by several plausible explanations independent of atrial fibrillation [25,26], and AC itself is a predictor of covert atrial fibrillation. Different markers of AC may contribute to stroke risk through distinct but overlapping pathways. Brain natriuretic peptide has been publicized for its relevance to heart failure and atrial fibrillation, primarily due to myocardial wall stretching [17,27]. LA dilation is also closely associated with atrial remodeling in response to multiple stressors [28]. A retrospective study of 32,454 community-dwelling individuals revealed that isolated LA enlargement without atrial fibrillation was a predictor of future stroke occurrence and development of atrial fibrillation [29]. Another possible mechanism for the increased risk of recurrent stroke is that functional and histological changes related to AC tend to cause embolic stroke. Pathological, structural, and functional changes in AC lead to aberrant blood flow and stasis in the atrial cavity, endothelial dysfunction, and hypercoagulability [30]. These characteristics predispose patients to thrombotic events. Ongoing research emphasizing unraveling the mechanisms underlying AC from histological to genetic perspectives, enables us to deeply associate their relevance to the current body of knowledge [31,32].

The tendency for severe neurological symptoms in patients with multiple AC factors is a significant finding. The severity of neurological deficits in patients with acute stroke varies according to the stroke subtype, with cardioembolic strokes often presenting more severe neurological deficits than other types [10]. Therefore, the higher NIHSS scores observed in the subgroup with multiple AC factors may suggest that an undisclosed etiology in this subgroup could be cardioembolic stroke.

Several limitations of this study must be considered when interpreting the results. First, generalizability was limited because all the subjects were Korean. Second, the retrospective design could have affected the outcome interpretation. To minimize the possibility of selection bias, PS-matching analysis was additionally employed, and the robustness increased in the PS-matched cohort. Third, the protocol for diagnosing ESUS may differ according to the participating site. The diagnosis was reviewed by a board-certified neurologist dedicated to stroke management. Fourth, the incomplete collection of AC factors hinders the precise estimation of effect sizes. The observed incompleteness may be attributable to the negative association of individual components, such as elevated NT-proBNP levels or LA enlargement, with vascular outcomes. However, both parameters showed higher values within the AC cohort, suggesting that the analysis may have been underpowered. In clinical practice, obtaining all possible biomarkers from patients is often impractical. Furthermore, the trend of results was reproduced in a subgroup analysis of 1,739 patients, who had complete data on AC factors (Supplementary Table 2), supporting the present data. Fifth, only LA enlargement and NT-proBNP levels were used as AC factors in this study. In this study, we excluded electrophysiological factors because of the difficulty in processing and ensuring uniformity across all centers. We recognize several limitations regarding our assessment of AC. First, our definition was confined to NT-proBNP elevation and LA enlargement, excluding electrophysiological markers such as P-wave terminal force. Second, NT-proBNP assays were selectively requested, primarily based on clinical suspicion of heart failure or radiographic cardiomegaly, rather than following a standardized protocol, thereby reducing the generalizability of our findings. Third, a significant proportion of missing NT-proBNP and echocardiographic data likely reduced the statistical power for detecting significant associations. Nevertheless, the independent association between NT-proBNP and recurrent stroke underscores its prognostic relevance, advocating for prospective and systematic biomarker assessment to enhance study robustness in future ESUS investigations.