TL;DR
This paper introduces SQUWA, a novel deep learning architecture that tackles the challenge of detecting Atrial Fibrillation (AF) from noisy photoplethysmography (PPG) signals. By integrating signal quality awareness directly into the learning process, SQUWA effectively utilizes partially corrupted data, achieving state-of-the-art performance and outperforming existing methods.
What Problem Does This Paper Address?
Atrial Fibrillation (AF) is a common heart arrhythmia that significantly increases the risk of stroke and other cardiovascular issues. Photoplethysmography (PPG) signals, commonly captured by wearable devices, offer a cost-effective and non-invasive way to monitor for AF. However, PPG signals are notoriously susceptible to noise and motion artifacts, especially in real-world, ambulatory settings. Traditional approaches often discard these noisy segments, which reduces the available data and can introduce biases, hindering accurate and continuous AF monitoring. This paper addresses the critical need for a method that can reliably detect AF even when PPG signals are compromised by noise.
Methods
The core innovation of SQUWA lies in its Signal Quality Aware design. Instead of treating noisy segments as unusable, SQUWA integrates an attention mechanism that dynamically assigns weights to different segments of the PPG signal based on their estimated quality. This allows the model to prioritize information from cleaner segments while still leveraging valuable data from partially corrupted ones.
Key components of the SQUWA architecture include:
- Composite Signal Generation: The raw PPG signal is combined with its first and second derivatives. This composite signal is then processed by variable-sized convolutional kernels to capture features at different temporal resolutions.
- CNN-LSTM Fusion: A Convolutional Neural Network (CNN) extracts features from the composite signal, and these features are then fed into a Long Short-Term Memory (LSTM) network for sequential learning.
- Signal Quality (SQ) Assessment: A pre-trained CNN-based model estimates the signal quality index (SQI) for each segment of the raw PPG signal. This SQI is derived using Class Activation Maps (CAMs).
- SQ-Attention Mechanism: This novel mechanism fuses the LSTM's hidden states with the SQI values. It learns to weigh the importance of different temporal segments based on their signal quality, effectively guiding the model's attention towards more reliable data.
Datasets
The study utilizes a comprehensive dataset strategy:
- Training Dataset: A large-scale dataset comprising 28,539 patients from hospital environments, with over 5.7 million PPG segments. This dataset is noted to have an estimated 25% label noise due to automatic monitor-based labeling.
- Signal Quality Dataset: A smaller dataset of 18,055 PPG segments from 13 patients, specifically curated with detailed annotations of signal quality, used to train the SQ assessment model.
- Test Datasets: Three external datasets were used for evaluation:
- Testset A (Public Source): From wrist-worn devices in ambulatory settings, containing 52,911 AF and 80,620 Non-AF samples from 163 patients.
- Testset B (Institution B): From Samsung Simband devices on 98 ambulatory patients, with 348 AF and 506 Non-AF segments.
- Testset C (Institution C): From fingertip PPG data of 126 hospital patients, containing 38,910 AF and 220,740 Non-AF segments.
Key Insights and Ideas
- Signal Quality is Paramount: The study strongly emphasizes that directly incorporating signal quality into the AF detection model is crucial for robust performance, especially in real-world scenarios with noisy data.
- Leveraging Corrupted Data: Instead of discarding noisy segments, SQUWA's attention mechanism intelligently utilizes them by down-weighting their influence, thereby maximizing the use of available data.
- Temporal Resolution Matters: The use of variable-sized convolutional kernels for composite signal generation allows the model to capture complex PPG patterns across different scales.
- Attention as a Quality Filter: The attention map visualizations clearly show that the model learns to focus on segments with higher signal quality, mirroring how human experts might interpret noisy signals.
- Generalizability: The model's strong performance across three independent test sets demonstrates its ability to generalize well to unseen data from different sources and conditions.
Key Results
SQUWA demonstrated superior performance compared to baseline models (ResNet-34, LSTM, ResNet-34+LSTM) across all evaluated metrics (AUROC, F1 score, AUCPR) on all three test sets.
- State-of-the-Art Performance: SQUWA achieved the highest AUCPR of 0.89 with label noise mitigation, surpassing even models trained with both ECG and PPG data (which achieved 0.86 AUCPR).
- Outperforming Existing Methods: SQUWA outperformed the CMC model, a recent method designed for AF detection with noisy labels, despite SQUWA not being explicitly designed for label noise mitigation.
- Ablation Study Highlights: The ablation study confirmed the importance of each component. Notably, removing the signal quality integration (NSQ variant) led to a significant performance drop, especially as signal quality worsened. The RSQ variant, using randomly generated SQIs, performed the worst, underscoring the critical role of accurate signal quality assessment.
Limitations and Future Work
- Not Fully End-to-End: The authors acknowledge that the signal quality assessment component requires pre-training on an annotated dataset, meaning the system is not entirely end-to-end.
- False Negatives: SQUWA can still be susceptible to inducing false negatives when artifacts obscure AF evidence.
- Future Work: The authors suggest that consecutive 30-second strips can be further analyzed to enhance model robustness for AF detection. They also propose that the principles of SQUWA's architecture could be adapted for other time-series data applications like human activity recognition and speech recognition. Further evaluations are needed to confirm its generalizability.
Summary updated