Atrial fibrillation detection from wearable photoplethysmography signals is highly affected by signal noise and motion artifacts. Most existing studies perform feature selection on pre-filtered datasets, where low-quality segments are removed using dataset-specific criteria. As a result, it remains unclear whether commonly used features are robust under varying signal quality conditions.
In this study, we investigate the effect of signal quality on feature-based atrial fibrillation detection using three datasets, including a public photoplethysmography dataset with long-term recordings from patients with suspected atrial fibrillation, a subset of an intensive care database, and a private clinical dataset of elderly high-risk patients. Signal quality is quantified using motion information derived from accelerometer signals, and data are stratified into multiple noise levels.
A total of 52 commonly used time-domain and variability features are extracted. Feature importance is evaluated under each noise level using model-based ranking methods. The top-ranked features are compared across noise conditions and datasets to assess the stability of feature selection. Classification performance is also evaluated to examine the impact of noise on model behavior.
Results show that feature importance varies across noise levels, with different features being ranked higher under different conditions. Features selected from higher-quality segments are not consistently retained as noise increases. In addition, the ranking of top features differs across datasets, indicating that feature selection is influenced by both signal quality and dataset characteristics. Classification performance shows a gradual decrease as noise level increases.
These findings suggest that feature selection for atrial fibrillation detection is sensitive to signal quality and may not generalize across datasets with different noise characteristics. Incorporating signal quality into feature design and model development may improve robustness in real-world wearable applications.