Detecting atrial fibrillation from wearable photoplethysmography signals is an important task in long-term cardiac monitoring, but model performance often depends on label availability and dataset characteristics. Pretrained models may help address this challenge by leveraging representations learned from large-scale physiological data. However, their usefulness for photoplethysmography-based atrial fibrillation detection has not been systematically evaluated.
In this study, we evaluate a pretrained cardiac sensing model on three datasets: a public dataset of 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. The pretrained model is fine-tuned on each dataset and compared with the same architecture trained from scratch.
Performance is assessed using standard classification metrics, and model behavior is further examined across datasets with different signal characteristics. Initial results suggest that the pretrained model provides more consistent performance across datasets than models trained from scratch. The advantage appears more noticeable on data with more complex signal characteristics, whereas differences are smaller under relatively favorable conditions.
These findings suggest that pretrained representations may improve generalization for photoplethysmography-based atrial fibrillation detection across diverse data sources. Further analysis will help clarify the conditions under which transfer learning is most beneficial for wearable cardiac monitoring.