TL;DR
This paper introduces a novel method for estimating instantaneous respiratory rate (IRR) from photoplethysmography (PPG) signals using the wavelet synchrosqueezed transform (WSST). The approach significantly outperforms traditional windowed methods, demonstrating high correlation with breath-by-breath variations and showing promise for continuous, non-invasive patient monitoring.
What Problem Does This Paper Address?
Continuous monitoring of vital signs is crucial for early detection of patient deterioration, especially in high-risk hospital settings. While respiratory rate (RR) is a key indicator, accurately and continuously estimating it from wearable sensors like PPG remains a challenge. Existing methods often rely on averaging RR over long time windows (e.g., 32 seconds), which can miss crucial rapid changes and subtle variations in breathing patterns. There's a need for methods that can capture instantaneous respiratory rate (IRR) with high fidelity from non-invasive, wearable PPG sensors.
Methods
The core of the proposed method lies in applying the wavelet synchrosqueezed transform (WSST) to extract respiratory information from PPG signals. Here's a breakdown:
- Respiratory-Induced Modulations: The paper first identifies three types of modulations in PPG signals caused by breathing: intensity, amplitude (AM), and frequency (FM). The study focuses on AM and FM.
- Signal Pre-processing: Raw PPG signals are first processed using Singular Spectrum Analysis (SSA). SSA is used here for denoising, removing baseline wander, and accurately detecting pulse peaks, which are essential for deriving the AM and FM modulations.
- Wavelet Synchrosqueezed Transform (WSST): This is the key novel technique. WSST is a time-frequency analysis method that enhances the spectral representation of a signal. It involves:
- Applying the Continuous Wavelet Transform (CWT) to the signal.
- Estimating instantaneous frequencies from the CWT coefficients.
- "Synchrosqueezing" the time-frequency representation to concentrate energy around the true instantaneous frequencies, effectively sharpening the spectral representation.
- IRR Estimation: The WSST is applied to the derived AM or FM modulations (resampled to 4 Hz). The instantaneous frequencies identified by WSST are then used to estimate the instantaneous respiratory rate.
- Masking: A crucial step involves masking the WSST spectrum using the averaged RR. This helps to focus the analysis on the expected respiratory frequency band, which can vary between individuals and over time.
- Comparison: The proposed IRR estimation is compared against:
- Averaged RR calculated using traditional 32-second windows (with and without overlap).
- Ground-truth data from a thermocouple placed under the nostrils.
Datasets
The study utilized two datasets:
- Custom Device Dataset: PPG signals were recorded from five healthy male subjects using a custom device with a reflectance PPG sensor. Data was collected from various body positions (forehead, ear, nose, neck, chest, arm, finger, and wrist). A thermocouple was used for ground-truth RR.
- Commercial Wrist-Worn Device Dataset: PPG signals were recorded from five healthy subjects (3 female, 2 male) using a commercial WaveletHealth wristband. A thermocouple was again used for ground-truth RR. This dataset included subjects with various skin colors.
Key Insights and Ideas
- Instantaneous vs. Averaged RR: The paper strongly advocates for the importance of instantaneous RR estimation. Averaging over long windows can obscure critical, rapid changes in breathing that are vital for clinical monitoring.
- WSST's Power for Breathing Signals: The WSST proves to be a powerful tool for extracting subtle frequency modulations in PPG signals that correspond to breathing. Its ability to sharpen time-frequency representations is key to isolating these signals.
- Body Position Matters: The location where the PPG sensor is placed significantly impacts the quality of the extracted respiratory signal. Upper body positions, particularly the chest, generally yielded better results than peripheral locations like the wrist or finger.
- AM vs. FM: While both AM and FM modulations were explored, the paper found FM to be a more reliable source for IRR estimation in their dataset.
- Masking for Robustness: Using averaged RR to mask the WSST spectrum is a clever way to improve the robustness of IRR estimation by focusing on the expected frequency range.
Key Results
- The proposed WSST-based method demonstrated a high correlation with breath-by-breath variations in the reference RR.
- The method produced more accurate results compared to traditional averaged RR methods using 32-second windows.
- For a set of five healthy subjects, the averaged similarity (measured by the Longest Common Subsequence - LCSS algorithm) between reference RR and the proposed instantaneous RR was 0.69. This significantly outperformed the similarity of 0.49 achieved with 32-second windowed averaged RR.
- Chest position consistently ranked among the top five locations for accurate IRR estimation across subjects.
- Upper body positions generally produced higher quality PPG signals for respiration extraction.
Limitations and Future Work
- Body Position Specificity: While various body positions were tested, a comprehensive comparison of all positions was outside the scope. The study identified promising locations but did not definitively declare a single "best" position for all scenarios.
- Motion Artifacts: The paper acknowledges that motion artifacts can affect PPG signals, especially at peripheral locations like the wrist. Future work could incorporate motion data (e.g., from accelerometers) to further improve robustness.
- Patient Data: The study was conducted on healthy subjects. Future work should involve long-term recordings from actual patients to evaluate the method's reliability in clinical settings.
- Reference Data Reliability: For long-term patient monitoring, ensuring the reliability of the ground-truth reference data is a challenge. The authors suggest exploring other validated sensors like wearable ECG patches.
- Filtering and Denoising: While WSST is powerful, further frequency-tracking filtering techniques (like Kalman smoothing) could be applied to remove sparse invalid IRR estimates and improve performance.
- IM Modulation: The study did not extensively explore Intensity Modulation (IM), noting that preliminary implementations did not yield reliable results. This could be a direction for future research.
Summary updated