Clinical performance of an atrial fibrillation burden tracking algorithm: Evaluation against reference public and clinical textile electrocardiographic datasets.

Background: Atrial fibrillation (AF), the most common sustained cardiac arrhythmia, increases risks of stroke, heart failure, and mortality. Short-term electrocardiographic (ECG) monitoring often misses paroxysmal or asymptomatic AF, underscoring the value of textile ECG platforms for continuous real-world rhythm assessment. Objective: To develop and clinically validate a lightweight, interpretable algorithm for AF detection and burden estimation using textile ECG recordings acquired during daily life. Methods: We developed Textile AF-Tracker (TAF-Tracker), an RR-interval-based machine learning pipeline using entropy, Lorenz-plot, statistical, and fragmentation features with ECG quality metrics to detect AF in 60-beat segments. Multiple classifiers (Random Forest, XGBoost, Support Vector Machine, logistic regression, and a threshold-based method) were trained on public AF datasets and long-term 3-lead textile ECG recordings (SKIIN™). Data were split by subject into training, validation, and test sets to ensure unseen data were tested. Results: Across public long-term AF datasets, Random Forest and XGBoost achieved 96%-99% accuracy, 95%-99% sensitivity, and 96%-99% specificity. In 14-day textile ECG recordings from 47 AF patients, XGBoost reached 98.4% accuracy (sensitivity 96.1%, specificity 98.7%). AF burden showed a median absolute error of 1.3% (interquartile range 0.7%-1.8%). In healthy and noise stress data, specificity remained ≥99%, even during activity. Conclusion: A lightweight RR-interval-based machine learning algorithm on textile ECG accurately detects AF and quantifies burden in long-term recordings with minimal error. Combined with a comfortable multi-lead textile platform, it provides a practical alternative to Holter monitors and implantable devices for continuous AF surveillance and treatment assessment.