Prediction of underlying atrial fibrillation in patients with a cryptogenic stroke: results from the NOR-FIB Study

The NOR-FIB study was an international prospective observational multicenter study designed to detect and quantify the burden of AF (≥ 2 min duration) in patients with CS or cryptogenic transient ischaemic attack (TIA) using an ICM (Reveal LINQ^®) and to identify biomarkers of incident AF [8]. CS was defined as a non-lacunar brain infarct in the absence of extra- or intracranial atherosclerosis (≥ 50% luminal stenosis in arteries supplying the ischaemic area), high-risk cardiac source (including patent foramen ovale) and any other specific cause of stroke. To avoid bias, clinical TIA cases were included only if an acute lesion on magnetic resonance imaging was detected. The pre-enrolment evaluation and CS diagnosis were assessed by the treating physician. All patients underwent 12-lead ECG and minimum 24-h rhythm monitoring prior to enrolment. Transthoracic echocardiogram was mandatory, while a transesophageal echocardiogram was requested in patients’ ≤ 65 years. Measurements were done according to current guidelines [9‐11]. Clinical data on vascular and AF risk factors according to predefined case report forms, and blood samples for biomarkers analyses were collected at enrolment and at a 12-month follow-up visit. All patients were monitored with the ICM for 12 months for arrhythmia detection [12].

Between January 2017 and September 2020, 259 finally included patients with CS or cryptogenic TIA from 18 participating centers in Norway, Denmark, and Sweden were monitored. Systematic ICM data evaluation verified paroxysmal AF or atrial flutter (≥ 2 min duration) in 74 (28.6%) patients. Results regarding arrhythmia detection are previously published [13]. For the purpose of AF prediction, differences between several clinical and paraclinical biomarkers in AF vs non-AF patients and the utility of eight existing AF-predicting scores were evaluated. Blood biomarker analyses were performed separately, showing significantly higher levels of cardiac biomarkers of which N-terminal pro-brain natriuretic peptide (NT-proBNP) was the strongest predictor of underlying AF (OR 4.8 [95% CI 1.8–13.0] in age and sex-adjusted model) [14].

Eligible risk scores tested in the NOR-FIB study, chosen due to data availability for each of the score components were as follows (presented predictive values according to original data*):

*The utility of the screening test is usually best expressed by its sensitivity (the percentage of true positive results) and NPV (correctly excluding individuals with no disease), and the AUC or c-index value as the measure of performance.

Cut-off values for the scores were chosen according to the optimal predictive performance from the original papers (AS5F, CHASE-LESS, STAF, SURF), a recommendation from the European Society of Cardiology (Brown ESUS-AF and HAVOC) [25], and comparison between the scores (CHA₂DS₂-VASc and HATCH) [15, 26].

Statistical evaluation was performed using IBM SPSS Statistics 28 software. Data were censored at the study exit, time of death or 12-month follow-up. Missing data on risk factors were registered as not present and included in the analysis. Patient characteristics and biomarker levels are presented as frequencies (%), mean (± standard deviation, SD) or median with interquartile range (IQR, Q1–Q3). Independent sample T-test or Mann–Whitney U-test, according to data distribution, was used to evaluate group differences for continuous variables and Pearson Chi-Square test or Fisher’s exact test for categorical variables. A p value < 0.05 was considered significant. Data analyses for arrhythmia detection and blood biomarker assessment are previously described [13, 14]. Binary logistic regression analyses were fitted to estimate the odds ratio (OR) together with a 95% confidence interval (CI). In a multivariate analysis (standard and stepwise), all relevant risk factors were included as covariates. For each patient, the eight AF-predicting scores were calculated and analyzed. The predictive performance of each score was assessed using receiver operating characteristic (ROC) curves, in addition to sensitivity, specificity, PPV and NPV. In the final evaluation, ROC analyses were performed simultaneously for comparison.

Risk stratification for underlying AF in stroke patients can be attempted by scrutinizing comorbidities predisposing to AF, biomarkers indicating atrial cardiopathy (the substrate for arrhythmia), and neuroimaging lesion patterns indicating embolic stroke [29‐33]. Even though increasing age is the most prominent risk factor, burden of comorbidities including hypertension, diabetes mellitus, heart failure, coronary artery disease, chronic kidney disease and obesity significantly contribute to AF development and progression [29]. Several ECG and echocardiographic findings increase the suspicion of atrial disease and AF [34‐37]. Elevated cardiac biomarkers levels [14, 38‐40] are also reported as a potentially surrogate marker for underlying paroxysmal AF and being incorporated in some of the predictive scores [41, 42]. The latest systematic review and meta-analysis on biomarkers for post-stroke AF detection assessed 69 multimodal markers and identified in all 26 clinical, ECG and blood-based biomarkers [7].

In the NOR-FIB study, several known clinical and paraclinical findings were associated with AF risk in univariate analysis (Table 2). Among the oldest CS patients (≥ 75 years) and those using antiarrhythmic drugs at admission, the likelihood of underlying AF was almost 50%. However, in multivariate analysis, the only independent predictors of AF were increasing age and smoking status. Our findings may be explained by the fact that several of the other AF risk factors are also closely linked, in different overlapping pathologies, to stroke in general. This is further the reason while a single biomarker, including cardiac natriuretic peptides should not be used as a decision-making tool [40], emphasizing the use of scores rather than a stand-alone approach.

Clinical risk scores may guide personalized evaluation approach in CS patients and contribute to more optimal access to key diagnostics. However, the criteria used to define patients at risk of underlying AF should be validated, easy to apply in the first days after admission, and commonly available in all units treating stroke patients. Scores with the highest sensitivity and highest NPV correctly identifying patients at the risk have the best utility. Recent systematic review on the use of risk scores for predicting new AF after stroke or TIA highlights seventeen different scores [41]. Age, LA size, hypertension, congestive heart failure, previous stroke/TIA and NIHSS levels are the throughout most common highly relevant components indicating underlying AF. All the scores are potentially eligible for applying during the diagnostic work-up of stroke patients, yet in more complex scores focus on proper clinical assessment covering the actual score components is essential. Five of these seventeen scores have been developed exclusively for CS or ESUS (embolic stroke of undetermined source) patients to predict AF: AF-ESUS, ACTEL, NDAF, Brown ESUS-AF and HAVOC [17, 18, 44‐46]. The first three scores were not applicable to test in the NOR-FIB study due to the use of different variables (LA measurements more often including volume rather than area, lack of specific data on non-stenotic plaques or tricuspid regurgitation). The last two scores, Brown ESUS-AF [16] and HAVOC score [21] showed lower sensitivity and NPV than the STAF and SURF, and lower AUC values than AS5F and SURF scores which predicted AF best. Even though the Brown-ESUS score is similar to the STAF score, the last also includes stroke severity (NIHSS) and may be used in combination with the D-dimer, increasing its accuracy (sensitivity 95%, specificity 100%) [47]. The more comprehensive HAVOC score, however, may be more relevant for younger or multimorbid CS patients’ population. Nevertheless, Brown ESUS-AF and HAVOC scores have lately been proposed by the European Society of Cardiology for the guidance of further prolonged ECG monitoring in CS [25]. In case of negative initial diagnostics (24 h ECG, echocardiography and hematological evaluation), if Brown ESUS-AF ≥ 2 or HAVOC score ≥ 4 prolonged ECG monitoring with external devices up to first 30 days or ICMs directly is recommended. Recently another AF-predicting score for CS patients monitored by ICM has been proposed, PROACTIA score based on premature atrial beats (PAC/24t), P-wave duration, P-wave morphology, and LA end-systolic volume index [48]. The score enables the identification of patients with low, intermediate and a high risk of subsequent AF detection (AUC 0.79), yet has a complicated formula for which the app-based solution is needed (at the moment unavailable) limiting its usability. Another new score stratifying AF risk in CS, the Graz AF Risk Score includes age, NT-proBNP, supraventricular premature beats, atrial runs, atrial enlargement, left ventricular ejection fraction and brain imaging markers [42]. This score was not applicable for us to test either, as we did not count supraventricular premature beats or atrial runs systematically. The usability of both the newest scores may be limited in stroke units mostly applying in-hospital telemetry instead of Holter monitoring, where atrial runs or premature beats counting is less used in the real-time evaluation assessment.

Initial work-up of CS requires wide radiological expertise and good quality of the cardiac and vascular evaluation to determine the probable stroke mechanism. Cardiac investigation focusing on the detection of cardio-aortic sources other than AF (Fig. 2), as well as the atrial cardiopathy signs is crucial in diagnosing cardioembolic stroke [49]. Furthermore, clinical or radiological evidence of multiple vascular territory strokes, cortical or posterior circulation lesions, vessel occlusions and greater stroke severity in the absence of significant stenosis in the ipsilateral artery should also lead the suspicion of potentially underlying cardioembolism and AF, if other cardiac sources excluded [33]. In-hospital ECG monitoring for at least 48 h to rule out AF and reflection on other possible causes of brain ischaemia (i.e., genetic disorders and low-risk cardiac sources) should be completed before the stroke may be classified as cryptogenic [6, 50].

For the purpose of prevention, it is realistic to recommend ICM monitoring only to eligible CS patients where underlying AF will affect change in secondary prevention (including those with contraindications to OAC, due to the advent of LA appendage closure procedures). From a socioeconomic perspective, CS patients with short life expectancy (< 1–2 year) or high disability (modified Rankin Scale ≥ 4 due to index stroke or other medical condition with no prospect of recovery), in whom the risk of suffering even a devastating AF-related stroke is less relevant for the outcome should be excluded. Patients with the highest AF suspicion should start prolonged monitoring as early as possible, as recurrent stroke often occurs within the first weeks. The best risk stratification approach is probably using scores combining age (OR 3.26), markers of atrial cardiopathy (OR 2.12–7.79), NIHSS (OR 2.5) and cardiac natriuretic peptides (OR 13.73) which, according to current knowledge, are the strongest predictors of AF (7). The scores that performed best in the present study, the AS5F, STAF and SURF highlight these biomarkers. In a recent study validating AF-predicting scores, the AS5F score demonstrated adequate discrimination (c-index 0.730), being useful in selecting patients for invasive arrhythmia monitoring [27]. The usefulness of the STAF and SURF score is due to its high NPV correctly identifying patients with the lowest AF-risk.

Our study has some minor limitations. First, functional and structural changes in the LA constituting the substrate for AF might have been even more frequent. As completion of the specified case report form for echo data was optional, some of the valuable information on atrial size was missing in 32 patients. Furthermore, the negative impact of hypertension on AF risk might have been leveled by the usage of antihypertensive drugs [43]. At last, there was no formal test of the performance of the evaluated clinical risk scores, however, in simultaneous ROC analysis scores’ AUC were similar to the AUC for continuous variables of each test. The best performance of the STAF, SURF and AS5F score may be due to a high proportion of patients ≥ 60 years (near 70%), yet the proportion ≥ 70 years was 40%, and ≥ 80 years 12%. However, as increasing age has shown to be one of the strongest predictors of AF in several similar studies evaluating biomarkers, and the risk factor that predicted AF among our patients, the findings are not surprising.