Abstract
Atrial fibrillation (AF), a common cardiac arrhythmia, significantly increases the risk of stroke, heart disease, and mortality. Photoplethysmography (PPG) offers a promising solution for continuous AF monitoring due to its cost efficiency and integration into wearable devices. Nonetheless, PPG signals are susceptible to corruption from motion artifacts and other factors often encountered in ambulatory settings. Conventional approaches typically discard corrupted segments or attempt to reconstruct original signals, reducing dataset size and introducing biases. We propose a novel deep learning model, Signal Quality Weighted Fusion of Attentional Convolution and Recurrent Neural Network (SQUWA), designed to retain accurate predictions from partially corrupted PPG. SQUWA integrates an attention mechanism that directly considers signal quality, dynamically adjusting the weights of time series segments. This approach enhances the influence of higher-quality segments while reducing that of lower-quality ones, enabling the utilization of a broader range of data. Our extensive experiments show that SQUWA outperforms existing PPG-based models, achieving a highest AUCPR of 0.89, exceeding models trained using both electrocardiogram (ECG) and PPG data.
Introduction
Atrial Fibrillation (AF) is a leading cause of stroke and heart disease globally. While Photoplethysmography (PPG)—the light-based technology found in most smartwatches—allows for continuous, non-invasive monitoring, it is notoriously sensitive to motion artifacts. In ambulatory settings, arm movement can easily corrupt the signal, making reliable detection difficult.
Traditional machine learning approaches handle this by either discarding "bad" data or attempting to reconstruct it. Discarding data creates monitoring gaps and can introduce selection biases, while reconstruction is computationally expensive and may introduce artifacts leading to false positives. SQUWA addresses these challenges by utilizing the entire signal stream, even when partially corrupted, without compromising diagnostic accuracy.
Methods
The SQUWA architecture consists of three innovative stages designed to maximize information extraction from noisy signals:
- Composite Signal Generation: The model processes raw PPG along with its first and second derivatives (representing blood flow velocity and acceleration). An "Attention Subnet" uses variable-sized convolutional kernels to capture patterns across different temporal resolutions.
- Feature Extraction (CNN-LSTM): A ResNet-based backbone extracts spatial features, which are then processed by a Long Short-Term Memory (LSTM) network to capture the temporal rhythm of the heart.
- Signal Quality (SQ) Attention: This core innovation uses a pre-trained signal quality model to generate a Signal Quality Index (SQI) over time. This SQI acts as a weight for the attention mechanism; noisy segments are down-weighted, ensuring the final AF classification is driven primarily by the clean portions of the data.
Datasets and Evaluation
The study utilized a diverse range of datasets to ensure robustness:
- Training Set: Over 5 million PPG samples from 28,539 hospital patients, with labels from bedside monitors.
- Signal Quality Dataset: 18,055 segments used to train the internal quality assessment component.
- Testset A (Public): Wrist-worn device data from ambulatory settings.
- Testset B (Institution B): Data from 98 patients using Samsung Simband wearables.
- Testset C (Institution C): Fingertip PPG data from 126 hospital patients, annotated by cardiac electrophysiologists.
Key Results and Insights
SQUWA demonstrated superior performance across all metrics and test sets. Notably, it achieved an AUCPR of 0.89 on Testset B, significantly outperforming standard ResNet and LSTM baselines. Remarkably, SQUWA (using only PPG) outperformed SiamAF, a state-of-the-art model trained on both PPG and ECG data (0.86 AUCPR).
Key insights from the study include:
- Soft Filtering: Weighting segments rather than deleting them maintains a continuous monitoring flow and preserves valuable information in partially noisy data.
- Physics-Informed Features: Including signal derivatives helps the model distinguish between irregular heart rhythms and motion noise.
- Label Noise Resilience: SQUWA proved highly resilient to the ~25% label noise present in the training data, outperforming specialized label-noise mitigation models like CMC.
Conclusion
SQUWA represents a significant advancement in wearable-based AF monitoring. By explicitly accounting for signal quality through an attention mechanism, it avoids the pitfalls of data truncation. While future work aims to make the system fully end-to-end and validate it across more diverse skin tones and activity levels, SQUWA currently provides a more accurate and less disruptive method for estimating AF burden in real-world conditions.
Summary updated