Predicting Incident Atrial Fibrillation after Stroke: A Scoping Review of Clinical Scores, Biomarkers, and AI-Enhanced Strategies

Figures

• 

  Fig. 1.. Predictive variables across atrial fibrillation risk scores and AI models. This heatmap displays the presence (blue) or absence (white) of the 25 most common predictive variables used across published AF risk scores and AI models. Each row is a model, and each column is a standardized clinical feature. Frequently used predictors include age, hypertension, heart failure, sex, BMI/obesity, CAD, diabetes, prior stroke/TIA, and ECG features. Advanced models add biomarkers (BNP/NT-proBNP), MRI findings, LVEF, EHR-based labs (creatinine, HbA1c), and vital signs. Additional factors include smoking, alcohol use, renal disease, cardiac murmur, PPG data, sleep apnea, and echocardiographic parameters. Overall, the heatmap highlights convergence in predictor selection across traditional and AI-based tools, emphasizing key domains for AF risk stratification [7, 10, 11, 13–17, 19, 20, 22–25, 27–34, 54–61].

• 

  Fig. 2.. Comparative performance of AFib prediction models. Comparative performance of published atrial fibrillation (AF) risk scores and artificial intelligence (AI)-based models. Bars represent the reported area under the curve (AUC) for each score or model, with error bars denoting the 95% confidence intervals where available. Classical scores (e.g., CHARGE-AF, FHS-AF) are shown alongside newer post-stroke-specific tools (e.g., BRAFIL, STAF) and modern AI systems (e.g., DeepECG-AI, UK Biobank AI, Khurshid et al. [15]) [10, 13–20, 22, 23, 41, 53–55, 58–62]. AF, atrial fibrillation; AI, artificial intelligence; AUC, area under the curve; ECG, electrocardiogram; MRI, magnetic resonance imaging.

• 

  Fig. 3.. Clinical variables used in AI models for AFib prediction after stroke, separated by model type legend [20, 54, 57, 63, 64]. This annotated heatmap shows clinical and imaging features used in AI models predicting incident AF after ischemic stroke. Each row corresponds to a study and its AI method. Green cells mark included features, and horizontal lines separate model categories. AF, atrial fibrillation; TIA, transient ischemic attack; ECG, electrocardiogram; HRV, heart rate variability; MRI, magnetic resonance imaging; LV, left ventricle; ML, machine learning.

• 

  Fig. 4.. Scatterplot of estimated AUC performance versus implementation cost (USD) for atrial fibrillation (AF) monitoring strategies. This scatterplot compares five evidence-based AF detection strategies by diagnostic accuracy (AUC), cost, and feasibility in low-resource settings. Risk-score screening (ASAS) has an AUC of 0.79 and is nearly cost-free, relying on clinical and echocardiographic data. Opportunistic Pulse Check from the SAFE trial shows an AUC of ∼0.90 at an estimated USD 5 per screening. Handheld ECG devices (e.g., Kardia) reach an AUC of 0.88 and cost about USD 100. Smartphone PPG apps (Apple, Huawei studies) show lower accuracy (AUC ∼0.63) and cost ∼USD 60/year. AI-enhanced Patch Monitors (e.g., Zio Patch) achieve an AUC of 0.87, providing 14-day continuous ECG monitoring for ∼USD 350. Circle size reflects estimated scalability and accessibility across health systems.