Atrial fibrillation after ischemic stroke: What risk scores illuminate and what they miss Authors: Silvio Saraullo, Andrea Palermi, Giulia Renda Published: 2026 We read with interest the study by Hsu et al. describing the Prediction of Atrial Fibrillation in Ischemic Stroke (PAFIS) score for risk stratification of atrial fibrillation detected after stroke (AFDAS) [1]. The authors address a practical question in stroke pathways: after ischemic stroke without previously recognized atrial fibrillation (AF), who deserves intensified rhythm surveillance? The issue is not merely diagnostic. Detecting AF may reclassify stroke mechanism and change secondary prevention when documented, whereas empirical anticoagulation based only on atrial cardiopathy has not improved outcomes in ARCADIA [2]. These are distinct clinical questions: PAFIS was developed to predict delayed AF documentation, not to establish stroke causality or anticoagulation benefit. Conversely, stroke-risk scores such as CHA₂DS₂-VASc and CHA₂DS₂-VA guide thromboembolic risk assessment once AF is documented, and treatment gaps persist even in high-risk AF patients [3]. Post-stroke AF scores should therefore guide surveillance intensity, not act as causal or therapeutic shortcuts. PAFIS assigns points for age ≥ 75 years, female sex, valvular heart disease, left atrial diameter ≥ 40 mm, and tricuspid regurgitation peak gradient ≥30 mmHg. Risk groups were derived from the total score and observed AFDAS gradient in the development cohort: low risk 0–1, intermediate risk 2–3, and high risk ≥4. In temporal validation, AFDAS was documented in 6.8%, 8.5%, and 14.0%, respectively, although the intermediate-risk group did not separate significantly from the low-risk group [1]. PAFIS is best viewed as an ordinal triage tool for expected monitoring yield, rather than as an absolute-risk calculator. Its main strength is the AFDAS-specific derivation in a broad ischemic stroke cohort rather than in cryptogenic stroke or embolic stroke of undetermined source (ESUS). Its variables capture a plausible phenotype of atrial and cardiopulmonary remodeling. However, “known AF” included AF documented before stroke or within 14 days after stroke, while AFDAS was restricted to AF first recorded thereafter [1]. This distinction matters because AFDAS is heterogeneous [4], [5], [6]. In some patients, AF was present before stroke but remained undiagnosed and may have caused the embolic event. In others, AF may be documented after acute brain injury as part of the stroke–heart syndrome. The 14-day threshold used by Hsu et al. should therefore be interpreted as an operational boundary based on timing of documentation, not as a validated mechanistic divider. Early AF may reflect previously undiagnosed AF, stroke-heart syndrome, or both; conversely, AF documented beyond 14 days cannot automatically be considered independent of the acute stroke-related process. Monitoring intensity shapes the endpoint. AF ascertainment in PAFIS relied on diagnostic codes and routine 12‑lead electrocardiography without standardized prolonged monitoring [1]. The score therefore predicts routine-care AF documentation, not true AFDAS incidence or burden, and may be influenced by differential opportunities for detection. Thus, observed AFDAS prevalence and meaning are inseparable from how, when, and in whom we look for it. Individual participant data from the ANTARCTICA pooled analysis confirmed that the interpretation of device-detected AF depends on the population and timing [7]. The derivation strategy raises methodological questions. Median follow-up differed markedly between development and validation cohorts, yet logistic regression treated patients without documented AF as non-events regardless of observation time. This may have affected discrimination and calibration by creating unequal opportunities for AF detection. The inclusion of clinically plausible predictors and the use of imputation are not limitations per se. However, the precision and stability of the coefficients, the clinical rationale and method used for variable selection, the extent and pattern of missing echocardiographic data, and imputation assumptions remain important for model transportability. In addition, converting model predictors into an integer point score and grouping patients into low-, intermediate-, and high-risk categories require prospective calibration against standardized monitoring strategies. Moreover, in temporal validation, the intermediate-risk group did not clearly separate from the low-risk group, and AF was still detected in 6.8% of low-risk patients. Decision curve analysis supports potential utility but does not prove that a PAFIS-guided pathway improves management. These aspects, together with calibration across settings, require sensitivity analyses and geographically distinct external validation. Existing post-stroke AF scores provide useful context for PAFIS but are not interchangeable. They capture different faces of post-stroke AF. In their original studies, Brown ESUS-AF, HAVOC, AF-ESUS, CHASE-LESS, and AS5F reported AUCs or c-indices ranging from approximately 0.725 to 0.848 [8], [9], [10], [11], [12], [13]. However, these estimates are not directly comparable because populations, endpoints, prediction horizons, and monitoring strategies differed [12]. Brown ESUS-AF is ESUS-specific; HAVOC is comorbidity-based; AF-ESUS is mainly useful for identifying patients at relatively low risk; CHASE-LESS is dependent on administrative and routine-care AF ascertainment; and AS5F is practical for early Holter-based selection. By contrast, CHA₂DS₂-VASc and CHA₂DS₂-VA were developed for thromboembolic risk stratification in documented AF, although CHA₂DS₂-VASc predicts new-onset AF and adverse outcomes without known AF [14]. In the PAFIS cohort, where several scores were assessed against the same endpoint, discrimination was modest and closely clustered: PAFIS 0.72, CHA₂DS₂-VASc 0.67, HAVOC 0.66, Brown ESUS-AF 0.65, and AF-ESUS 0.62 [1]. The practical implication is a hypothesis-generating tiered surveillance framework, not a validated clinical algorithm. In principle, low-risk patients could continue with standard telemetry and routine follow-up, intermediate-risk patients could be considered for structured external monitoring, and high-risk patients for prolonged external monitoring or an implantable cardiac monitor (ICM) when feasible. However, these allocations remain speculative. PAFIS categories were not prospectively linked to predefined monitoring modalities, number needed to monitor, cost-effectiveness thresholds, or clinically established decision cut-offs. Thus, score-guided rhythm surveillance should be considered a testable care pathway and requires prospective evaluation before clinical implementation. Future studies should use prespecified prediction horizons, time-to-event methods, standardized monitoring across score strata, electrocardiographic and echocardiographic strain markers, and clinically relevant endpoints incorporating AF timing and burden. AI-enabled analysis of sinus-rhythm 12‑lead ECGs may add latent signatures of atrial vulnerability and complement clinical scores [15], [16], but requires prospective post-stroke validation. Taken together, available post-stroke AF scores, including PAFIS, should be interpreted according to the clinical question they were designed to address: enrichment for AF detection. They can help clinicians decide where rhythm surveillance should be intensified, but they should not be used to infer whether AF caused the index stroke or to guide anticoagulation in the absence of documented AF (Fig. 1). The next step is to test whether score-guided monitoring improves diagnosis, management, and outcomes. Comparative risk scores for atrial fibrillation detection after ischemic stroke. PAFIS is shown as the principal score discussed in this Correspondence, together with its components, risk categories, and development and temporal validation performance. The ROC panels are adapted from the discrimination analyses reported in the PAFIS study and are shown in simplified graphical form. Comparator scores are provided for clinical context only, with AUCs derived from their original derivation and/or validation studies. The proposed rhythm-surveillance framework shown in the lower panel is hypothesis-generating and requires prospective validation before clinical implementation. AF, atrial fibrillation; AFDAS, atrial fibrillation detected after stroke; AIS, acute ischemic stroke; AUC, area under the curve; BMI, body mass index; CAD, coronary artery disease; CHF, congestive heart failure; ECG, electrocardiogram; ESUS, embolic stroke of undetermined source; HTN, hypertension; ICM, implantable cardiac monitor; LA, left atrium; LVEF, left ventricular ejection fraction; NIHSS, National Institutes of Health Stroke Scale; PVD, peripheral vascular disease; ROC, receiver operating characteristic; TIA, transient ischemic attack; TRPG, tricuspid regurgitation peak gradient. CRediT authorship contribution statement Silvio Saraullo: Writing – original draft. Andrea Palermi: Writing – original draft. Giulia Renda: Writing – review & editing, Supervision, Conceptualization. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. References