Heart Rate Extraction from Abdominal Audio Signals

Jake Stuchbury-Wass, Erika Bondareva, Kayla-Jade Butkow, Sanja Scepanovic, Zoran Radivojevic, Cecilia Mascolo

Abstract

Researchers from the University of Cambridge and Nokia Bell Labs have demonstrated the feasibility of accurately extracting heart rate (HR) from audio recorded at the abdomen—a region typically reserved for monitoring gastrointestinal (GI) health. By employing a specialized digital signal processing pipeline centered on wavelet denoising, the system achieves an average heart rate error of 3.4 BPM. This performance makes abdominal audio monitoring competitive with consumer-grade wrist-worn wearables, such as the Apple Watch, for resting heart rate tracking.

Introduction

Monitoring gastrointestinal health often involves a clinician using a stethoscope to listen to abdominal sounds (ABS), a process that provides only a brief snapshot of a patient's condition. While wearable belts are being developed for continuous GI monitoring, they are frequently limited to a single purpose.

The primary challenge in expanding the utility of these sensors is the high level of noise in the abdominal region. Heart sounds (the "lub-dub" of valves closing) must travel from the chest to the abdomen, arriving significantly attenuated. Once there, they are obscured by louder sounds from respiration, digestion (GI motility), and physical movement. This research investigates whether abdominal audio can be sufficiently cleaned to extract reliable heart rate data, effectively transforming a GI monitor into a multi-purpose vital sign tracker.

Methodology

To overcome the low Signal-to-Noise Ratio (SNR) of abdominal heart sounds, the authors developed a four-stage digital signal processing (DSP) pipeline:

  1. Pre-processing: Audio is downsampled to 4 kHz and segmented into 10-second windows. Z-score normalization is applied to identify and discard segments containing significant motion artifacts.
  2. Signal Denoising: Because heart and GI sounds overlap in frequency, standard filtering is insufficient. The pipeline utilizes a 5th-order Butterworth lowpass filter (200 Hz cutoff) followed by Wavelet Denoising. Specifically, a Coiflet 4 wavelet with a 5th-level decomposition and soft thresholding is used to isolate rhythmic heart peaks from transient noise.
  3. Heart Rate Estimation: Following denoising, the system performs peak detection. To minimize false positives from digestive sounds, a minimum inter-peak distance of 0.65s is enforced, capping the detectable resting heart rate at 92 BPM.
  4. Post-processing: The system calculates both instantaneous HR (iHR) and average HR (aHR), applying moving average filters to smooth the final output.
The multi-stage digital signal processing pipeline designed to extract heart rate from noisy abdominal audio data.

Experimental Setup

The study utilized a dataset comprising 104 hours of audio collected from seven participants over a 10-day period. Data was captured using a custom-built electronic stethoscope embedded in a wearable belt. For ground truth comparison, participants wore a Polar H9 ECG chest strap, providing a gold-standard electrical heart rate signal. Notably, data collection occurred remotely during the COVID-19 pandemic, ensuring the results reflect real-world variability in device placement and user activity.

Results and Discussion

The system's performance was evaluated using Mean Absolute Error (MAE) and Mean Deviation Error (MDE). The average heart rate (aHR) achieved an MAE of 3.4 BPM (4.8% error), while the instantaneous heart rate (iHR) showed an MAE of 6.4 BPM (8.9% error).

Performance metrics (MDE and MAE) for average heart rate (aHR) and instantaneous heart rate (iHR) across the seven study participants.

When benchmarked against established wearables from prior studies, the abdominal audio method (3.4 BPM MAE) outperformed the Empatica E4 (11.3 BPM) and proved highly competitive with the Apple Watch 4 (4.4 BPM) and Fitbit Charge 2 (7.3 BPM) for resting heart rate monitoring.

Key Insights

  • Sensor Fusion: The ability to extract HR from abdominal sensors allows for a "two-for-one" monitoring solution, reducing device fatigue for patients who already require GI monitoring.
  • The Power of Wavelets: Standard band-pass filtering fails because heart sounds and stomach gurgles share similar frequency bands. Wavelet transforms allow the algorithm to analyze the signal in both time and frequency domains, which is essential for isolating the heart's signature.
  • Resting State Constraints: The current algorithm is optimized for resting heart rates (50–90 BPM). While this increases robustness against noise, it introduces a bias that pulls estimates toward a central average, potentially limiting accuracy during physical exertion.

Conclusion and Future Work

While the results are promising, the study identified several limitations. Approximately 10.2% of the data was discarded due to excessive noise from movement or loud GI events. Furthermore, the heuristic-based peak detection limits the system to resting heart rates. Future research will focus on replacing these heuristics with machine learning models, such as LSTMs, to better segment heart sounds from background noise and to extract more complex metrics like Heart Rate Variability (HRV) and respiratory rate.

Summary updated

BibTeX
@inproceedings{stuchburywass2023heart,
  author = {Jake Stuchbury-Wass and Erika Bondareva and Kayla-Jade Butkow and Sanja Scepanovic and Zoran Radivojevic and Cecilia Mascolo},
  title = {{Heart Rate Extraction from Abdominal Audio Signals}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2023},
  doi = {10.1109/icassp49357.2023.10096600},
}
Figure