TL;DR
This paper introduces a novel machine learning approach for estimating breathing rate using photoplethysmography (PPG) data from a head-worn VR mask. The method achieves high accuracy, outperforming existing techniques, and demonstrates the feasibility of non-intrusive breathing rate monitoring in real-world scenarios.
What Problem Does This Paper Address?
Breathing rate is a crucial vital sign, but accurately measuring it outside of controlled laboratory settings, especially with novel wearable devices, remains a challenge. While heart activity is easier to monitor, deriving breathing rate from it is complex. Existing methods often struggle with motion artifacts and the unique sensor placement of new devices like head-worn VR masks. This paper aims to bridge this gap by developing a robust and accurate breathing rate estimation algorithm specifically for PPG data from a forehead-mounted sensor.
Methods
The core of the proposed method is a machine learning pipeline built upon advanced signal processing. Key steps include:
- Data Segmentation: Raw PPG data is segmented into 20-second windows, a shorter duration than many previous methods, allowing for quicker response to physiological changes.
- Dual Filtering: Two distinct filtering procedures are applied. The first prepares data for P-P (peak-to-peak) interval extraction, while the second focuses on isolating breathing-related frequencies for feature extraction.
- Motion Artifact Removal and Quality Assessment: A novel procedure using accelerometer data quantifies movement and identifies corrupted PPG segments, ensuring that only reliable data is used for analysis. This is crucial for wearable devices that are prone to motion.
- Feature Extraction: Four categories of features are extracted from the processed PPG data and P-P intervals: PRV (Pulse Rate Variability)-based, RR-PSD (Respiration Rate Power Spectral Density)-based, peak-valley-based, and PPG-PSD-based. This multi-faceted approach aims to capture various aspects of breathing-induced modulations in the PPG signal.
- Machine Learning Model: An Extreme Gradient Boosting (XGBoost) regression model is trained on these features to predict the breathing rate.
Datasets
The study utilized two datasets collected from a total of 37 healthy subjects wearing a head-mounted VR mask with an integrated PPG sensor.
- Dataset 1: 27 subjects (17 male, 10 female) performed controlled breathing exercises with varying breath durations (3s, 5s, 8s).
- Dataset 2: 13 subjects (7 male, 6 female) followed audio cues for breathing synchronization at specific rates.
The researchers provide public access to one of these datasets (13 subjects) as a benchmark for future research.
Key Insights and Ideas
- Shorter windows are better: Contrary to some previous work, this study found that a shorter window size (20s) for PPG data processing leads to better performance, making the algorithm more responsive and practical for real-time applications.
- Fusion of features is key: Combining features from different aspects of the PPG signal (PRV, PSD, peak-valley) significantly improves accuracy compared to using any single feature group alone. This highlights the complex relationship between breathing and PPG signals.
- Motion artifact removal is critical: The novel motion artifact detection and removal step, using accelerometer data, demonstrably improves the robustness and accuracy of breathing rate estimation, especially in scenarios with potential movement.
- Head-worn PPG is viable: The research successfully demonstrates that reliable breathing rate estimation is achievable using PPG sensors placed on the forehead, opening doors for more integrated and less intrusive wearable health monitoring.
Key Results
The proposed method achieved impressive results:
- Outperformed baselines: The algorithm consistently outperformed three baseline methods (Dominant Frequency, HeartPy, and a Deep Learning approach) across all datasets.
- Mean Absolute Error (MAE): An MAE of 1.38 breaths per minute was achieved when using the combined datasets and the full processing pipeline (including motion artifact removal).
- Pearson Correlation Coefficient (PCC): A PCC of 0.86 was achieved, indicating a strong linear correlation between the estimated and true breathing rates.
- Window size impact: The method performed best with a 20s window size, showing that shorter windows are advantageous for this specific approach.
- Feature importance: PPG-PSD features were found to be highly informative, but their fusion with other feature groups yielded the best overall results.
Limitations and Future Work
The authors acknowledge several limitations and propose future directions:
- Stationary Data: The datasets were collected in relatively stationary conditions. Future work will focus on developing and testing models with "in-the-wild" data, which is more challenging due to diverse motion patterns.
- Out-of-Distribution Breathing Rates: The current models were trained on data within the normal breathing rate range. Their performance on significantly higher or lower rates is an open question.
- Semi-supervised and Self-supervised Learning: To address the challenges of unlabeled data and human motion, the researchers plan to explore semi-supervised and self-supervised machine learning approaches.
- Synthetic Data: Generating synthetic data with varying motion levels is also proposed to create more robust models.
Summary updated