Rapid extraction of respiratory waveforms from photoplethysmography: A deep corr-encoder approach

Harry J. Davies, Danilo P. MandicBiomedical Signal Processing and Control 2023 · DOI

Abstract

Much of the information of breathing is contained within the photoplethysmography (PPG) signal, through changes in venous blood flow, heart rate and stroke volume. We aim to leverage this fact, by employing a novel deep learning framework which is a based on a repurposed convolutional autoencoder. Our model aims to encode all of the relevant respiratory information contained within photoplethysmography waveform, and decode it into a waveform that is similar to a gold standard respiratory reference. The model is employed on two photoplethysmography data sets, namely Capnobase and BIDMC. We show that the model is capable of producing respiratory waveforms that approach the gold standard, while in turn producing state of the art respiratory rate estimates. We also show that when it comes to capturing more advanced respiratory waveform characteristics such as duty cycle, our model is for the most part unsuccessful. A suggested reason for this, in light of a previous study on in-ear PPG, is that the respiratory variations in finger-PPG are far weaker compared with other recording locations. Importantly, our model can perform these waveform estimates in a fraction of a millisecond, giving it the capacity to produce over 6 hours of respiratory waveforms in a single second. Moreover, we attempt to interpret the behaviour of the kernel weights within the model, showing that in part our model intuitively selects different breathing frequencies. The model proposed in this work could help to improve the usefulness of consumer PPG-based wearables for medical applications, where detailed respiratory information is required.

Introduction

While wearable devices like smartwatches are ubiquitous, the clinical depth of the information they provide is often limited. Photoplethysmography (PPG)—the optical technology used to track heart rate—contains subtle signatures of respiration encoded through changes in blood volume and heart rate variability. Most existing research focuses on extracting a single metric: the Respiratory Rate (RR). However, the full respiratory waveform contains richer diagnostic data, such as inspiration-to-expiration ratios.

This study addresses the challenge of reconstructing full breathing waveforms from weak and noisy finger-PPG signals. By bridging the gap between simple wearable sensors and medical-grade monitoring like capnography, the researchers aim to enhance the utility of consumer health tech.

Methods

The core of the proposed framework is a 1D Convolutional Autoencoder. While autoencoders are typically used for unsupervised tasks like denoising, this architecture is repurposed for supervised waveform translation.

Architecture of the 1D Convolutional Autoencoder for respiratory waveform extraction.

The model consists of two main components:

  • The Encoder: Three convolutional layers (8 kernels each) compress a 9.6-second window of PPG data into a low-dimensional latent space, forcing the network to extract fundamental respiratory features.
  • The Decoder: Three transposed convolutional layers up-sample the latent representation to reconstruct the respiratory waveform.

The model was trained using Mean Squared Error (MSE) against "gold standard" capnography (CO2) references. Large kernel sizes (up to 150 samples) were utilized to capture the slow, rhythmic nature of human breathing.

Results and Discussion

The model demonstrated state-of-the-art performance across two benchmark datasets: Capnobase (42 subjects) and BIDMC (hospital patients).

Comparison of reconstructed waveforms and quantitative error analysis against Partial Least Squares (PLS).

Performance and Efficiency

On the Capnobase dataset, the model achieved a median Mean Absolute Error (MAE) of 0.37 breaths per minute, outperforming traditional statistical methods like Partial Least Squares (PLS) regression. Even when tested on the noisier BIDMC dataset, it maintained a median MAE of 0.89 bpm.

Beyond accuracy, the model is exceptionally efficient. It can process a 9.6-second window in just 0.43 milliseconds on a standard laptop CPU, enabling the generation of over 6 hours of respiratory data in a single second of compute time.

Interpretability

Analysis of the kernel weights revealed that the model acts as a "learned filter bank." Different kernels tuned themselves to specific breathing frequencies (e.g., slow deep breaths vs. rapid breathing), suggesting the model intuitively learns the underlying physiology rather than just memorizing patterns.

Limitations

A key limitation was the "Duty Cycle"—the precise timing of breath phases. The model struggled to perfectly match the inspiration/expiration ratio. This is likely due to the physiological limitations of finger-based PPG, where respiratory variations are significantly weaker than in other locations like the ear.

Conclusion

This deep encoder approach provides a rapid, accurate, and computationally lightweight method for extracting respiratory waveforms from PPG. Its efficiency makes it a prime candidate for integration into low-power wearable microcontrollers, potentially transforming consumer devices into more robust medical monitoring tools.

Summary updated

BibTeX
@article{davies2023rapid,
  author = {Harry J. Davies and Danilo P. Mandic},
  title = {{Rapid extraction of respiratory waveforms from photoplethysmography: A deep corr-encoder approach}},
  journal = {Biomedical Signal Processing and Control},
  year = {2023},
  doi = {10.1016/j.bspc.2023.104992},
}

Loading related content...

Figure