TL;DR
This paper introduces "Nightbeat," a novel method for estimating heart rate (HR) from wrist-worn accelerometers during sleep, even in uncontrolled home environments. They also release "Nightbeat-DB," a new dataset specifically designed for this task, which enables them to demonstrate their method significantly outperforms existing approaches, achieving medical-grade accuracy.
What Problem Does This Paper Address?
Large-scale health studies, like the UK Biobank, often collect extensive accelerometer data from participants to track activity and sleep patterns. However, these studies typically lack direct heart rate (HR) measurements. While modern smartwatches use optical sensors for HR, these are not always present in research settings. This gap means valuable insights into cardiovascular health and disease progression, which can be gleaned from HR dynamics during sleep, are missed. The challenge lies in accurately extracting HR from accelerometer data alone, especially given the presence of motion artifacts, which are common even during sleep. Previous attempts have often relied on controlled lab settings or struggled with real-world noise.
Methods
The core of the "Nightbeat" method involves several key steps:
- Nightbeat Signal Generation: The raw 3-axis accelerometer data is first filtered to isolate frequencies relevant to the ballistocardiogram (BCG) – the subtle vibrations caused by heartbeats. The L2-norm of these filtered axes is then computed, followed by further filtering to create the "Nightbeat signal."
- Motion Artifact Removal: Using Short-Time Fourier Transforms (STFT) of the Nightbeat signal, the authors identify segments with unusually high spectral energy, flagging them as motion artifacts. This is done by comparing the energy to a robust median and standard deviation.
- Frequency Peak Identification (Curve Tracing): In segments free of motion artifacts, the STFT is used to identify potential HR peaks in the frequency domain. These peaks are then "connected" across time to form "HR curves," allowing for gaps and jumps. This process is inspired by curve tracing methods used for photoplethysmography (PPG) signals.
- Heart Beat Detection: Based on the identified HR curve (HRC), a narrow frequency band is defined around the estimated HR. Within this band, peaks in the time-domain signal are detected to identify individual heartbeats and derive an inter-beat interval-based HR prediction (HRP).
- Post-Processing (Rolling Median): Finally, a rolling median of the frequency-based HR predictions (HRC) is computed over 5-minute windows. This HRM serves as a quality check, and the final prediction is made if the HRC and HRP are sufficiently close and within a reasonable range of the HRM.
Datasets
- Nightbeat-DB (New Dataset): This is the primary dataset introduced in the paper. It comprises synchronized wrist-worn accelerometer (Axivity AX3) and ECG chest belt (movisens EcgMove 4) recordings from 42 participants over 310 nights in uncontrolled at-home settings. Participants also provided questionnaires about their sleep habits, including whether they shared a bed.
- AW Dataset: A publicly available dataset from Physionet, consisting of wrist-worn accelerometer data from 31 participants recorded in a sleep laboratory. This dataset provides HR values derived from the Apple Watch's PPG sensor, which are used as ground truth (though linearly interpolated).
Key Insights and Ideas
- The Power of BCG during Sleep: The paper strongly validates the potential of using the ballistocardiogram (BCG) captured by accelerometers for HR estimation, particularly during sleep when voluntary motion is minimal.
- Uncontrolled Settings are Key: The introduction of Nightbeat-DB highlights the critical need for datasets that mimic real-world, uncontrolled environments, as opposed to solely lab-based recordings, to truly assess the robustness of wearable sensing methods.
- Motion Artifacts are the Primary Hurdle: The success of Nightbeat hinges on its effective motion artifact detection and removal, demonstrating that this is the most significant challenge in extracting reliable HR from accelerometer data.
- Bed-Sharing Matters (Especially for Women): The study found that sharing a bed significantly increased the mean absolute error (MAE) for female participants, suggesting that vibrations coupling through mattresses can introduce noise that is harder to filter out. This effect was less pronounced for male participants.
Key Results
- Superior Accuracy: Nightbeat achieved a Mean Absolute Error (MAE) of 0.88 bpm on the Nightbeat-DB dataset, which is 76% better than previous approaches. This performance is also noted to be better than established medical standards for heart rate monitors.
- High Correlation: The method achieved an average correlation of 0.81 across participants on Nightbeat-DB and 0.95 when considering all participants together.
- AW Dataset Performance: On the AW dataset, Nightbeat achieved an MAE of 1.68 bpm with an average correlation of 0.64.
- Baseline Performance: Existing methods like PWR and Jerks showed significantly higher MAEs (e.g., 8.83 bpm and 13.23 bpm on Nightbeat-DB), and BioInsights performed poorly with MAEs around 20 bpm. Even a simple "mean-prediction" baseline outperformed these methods.
- Bed-Sharing Impact: Female participants experienced a 32% increase in MAE when sharing a bed, while male participants saw a 2% decrease.
Limitations and Future Work
- Limited HR Range: The datasets used primarily contain HRs below 90 bpm. The authors acknowledge that evaluating the method on patients with higher HRs during sleep or those with cardiovascular diseases, which can lead to different HR patterns, is necessary.
- Abrupt HR Changes: Nightbeat may struggle to track very rapid HR changes (e.g., >10 bpm within minutes) due to its median smoothing in post-processing. While not observed in the current datasets, this could be an issue in patients with sleep disorders.
- Awake State: The current Nightbeat method is specifically designed for sleep and is unlikely to extend well to awake states due to the higher frequency and duration of movements, which would challenge the motion artifact removal and curve tracing components.
- Future Directions: The authors suggest evaluating Nightbeat's utility in disease modeling within existing large longitudinal studies and extending similar methods to estimate HR during awake states, which would require more robust motion artifact handling and faster HR tracking.
Summary updated