TL;DR
This paper introduces a novel, non-intrusive method for estimating heart rate (HR) from ballistocardiography (BCG) signals captured by sensors in chairs or mattresses. By combining Hilbert Transform to extract the BCG pulse envelope with a Phase Vocoder for enhanced frequency resolution, the method achieves highly accurate HR estimation, outperforming existing techniques.
What Problem Does This Paper Address?
Accurate and convenient heart rate monitoring is crucial for managing cardiovascular diseases. While wearable sensors exist, they can be uncomfortable. Ballistocardiography (BCG) offers a non-intrusive alternative, capturing the body's subtle movements caused by blood ejection. However, traditional methods for extracting HR from BCG signals often struggle with noise and irregular signal patterns, leading to inaccurate estimations. This paper aims to overcome these limitations by developing a more robust and precise HR estimation technique from BCG.
Methods
The proposed method, termed Hilbert Transform with Phase Vocoder (HTPV), involves two main stages:
- Hilbert Transform (HT): BCG signals are modeled as a combination of respiration, a modulated pulse envelope (carrying HR information), and noise. After removing respiration and noise using a band-pass filter, HT is applied to the remaining signal. HT extracts the pulse envelope by creating an analytical signal and then taking its magnitude squared. This envelope effectively models the periodic J-peaks in the BCG signal, which correspond to heartbeats.
- Phase Vocoder (PV): While the envelope's frequency spectrum contains HR information, its resolution is limited by the window size used in Fast Fourier Transform (FFT). To overcome this, a Phase Vocoder is employed. It analyzes phase changes between successive frames to precisely estimate instantaneous frequency, significantly improving frequency resolution. This allows for a more accurate determination of the fundamental frequency of the pulse envelope, which directly translates to the heart rate.
Datasets
The study utilized BCG and ECG signals recorded from seven human subjects. The BCG signals were captured using chair-based sensors, while ECG served as the ground truth, recorded from chest-based electrodes. All signals were sampled at 225 Hz. The data collection involved subjects sitting in a chair, with the freedom to change posture (e.g., leaning, bowing) to simulate real-world conditions. The recordings lasted approximately 15 minutes per subject. No new benchmark datasets were introduced; existing methods were compared against the proposed method using this collected data.
Key Insights and Ideas
- Envelope is Key: The core insight is that the pulse envelope of the BCG signal, rather than the raw signal itself, contains the most reliable HR information. Traditional methods often fail because the raw BCG signal is heavily influenced by noise and other physiological signals.
- Frequency Domain Power: By transforming the problem into the frequency domain using FFT and then refining it with the Phase Vocoder, the method leverages the periodic nature of the heart's activity more effectively than time-domain approaches.
- HT for Noise Resilience: The Hilbert Transform proves to be a powerful tool for isolating the HR-related envelope, making the system more resilient to the inherent noise and variability in BCG signals.
- PV for Precision: The Phase Vocoder's ability to enhance frequency resolution is critical for accurately pinpointing the HR, especially when dealing with signals that might have subtle variations or when the sampling rate isn't extremely high.
Key Results
The proposed HTPV method demonstrated excellent performance:
- Accuracy: Achieved a Mean Absolute Error (MAE) of 0.90 BPM and a Standard Deviation (STD) of 1.14 BPM.
- Correlation: A Pearson correlation coefficient of 0.98 was obtained between the estimated HR and the ground truth HR from ECG.
- Comparison: The HTPV method significantly outperformed existing methods like SEN [5], [8], and SLT [9] in terms of MAE and STD, as detailed in Table I. For instance, the mean MAE for HTPV was 0.90 BPM, compared to 1.57 BPM for SEN, 2.44 BPM for [8], and 1.57 BPM for SLT [9].
Limitations and Future Work
- Limitations: The paper acknowledges that the performance of the Phase Vocoder can be affected by noise. While the Hilbert Transform helps in extracting the envelope, significant noise might still impact the accuracy. The method's performance with extreme posture changes or vigorous movements was not extensively detailed.
- Future Work: The authors suggest that further improvements could be made by exploring more advanced noise reduction techniques. Additionally, investigating the method's robustness across a wider range of physiological conditions and movement artifacts would be beneficial for real-world deployment. The potential for integrating this method into practical, everyday devices for continuous monitoring is also a clear avenue for future development.
Summary updated