Aims: Photoplethysmography (PPG) allows the monitoring of patients' cardiac activity non-invasively and continuously using wearable devices. This approach could potentially enable early detection of cardiovascular diseases and critical pathological events. However, PPG signals acquired in daily life are often corrupted by movement artifacts. Effective denoising is challenging because the frequency bands of the artifacts and the signal often overlap. Existing traditional denoising methods underperform in the case of severe movement artifacts. Some published deep learning denoisers outperform traditional techniques but lack interpretability, which is crucial for troubleshooting and providing explicit information for clinical decision-making.
Methods: In this study, we use an unfolded deep neural network to denoise PPG signals. A dictionary of kernels is learned, which captures recurrent morphological patterns of the signal, improving the model's interpretability. A convolutional sparse coding model reconstructs each signal by selecting the optimal kernels from the dictionary and computing their temporal locations. The convolutional framework allows reusing the same kernels at different temporal locations and enables the processing of the whole signal without the need for pulse segmentation. The network is trained and tested with PPG signals from the PulseDB database and corrupted with a synthetic movement artifact model from the literature. The method's performances are compared with a reference convolutional denoising autoencoder.
Results: The proposed method achieved a higher median signal-to-noise ratio than the reference method (9.97 dB vs 6.27 dB) and a smaller median mean absolute error between the heart rate of the target and reconstructed PPG signals (4.49 bpm vs 7.26 bpm for proposed and reference methods, respectively).
Conclusion: Our approach promises to be a reliable strategy for improving the quality of PPG signals from wearable devices. Furthermore, its potential extends beyond denoising, as it can also extract meaningful waveform features, which may reveal ongoing cardiovascular alterations.