BallistoBud: Heart Rate Variability Monitoring using Earbud Accelerometry for Stress Assessment

Research Background and Issues

• Issues and Challenges: With the widespread adoption of earbud devices, the authors identified challenges in using sensors such as inertial measurement units (IMU) and photoplethysmography (PPG) to monitor physiological data for psychological stress assessment. Current earbud technologies face data quality issues, particularly when using ballistocardiogram (BCG) signals generated by IMU, which are influenced by individual differences, earbud positioning, and body movements.
• Research Significance: Accurate heart rate (HR) and heart rate variability (HRV) are critical physiological risk indicators for stress levels. If low-power, non-invasive physiological monitoring can be achieved through earbud devices, it could significantly advance the development of comfortable and portable mental health management technologies.
• Related Research and Motivation: Research on earbud devices is currently limited, with most studies focusing on PPG-based heart rate monitoring. Although IMU sensors are commonly integrated into earbuds, their potential for heart rate and psychological stress monitoring remains underexplored. Inspired by previous work (e.g., using BCG to extract HRV), this study aims to systematically evaluate the performance of IMU and PPG and propose new methods to improve BCG data quality.

Solution

• Methods and Innovations:
  • A machine learning-based algorithm framework is proposed to evaluate and filter low-quality IMU data, enhancing the robustness and accuracy of psychological stress monitoring using BCG signals.
  • A novel visualization technique ("ECG-gated BCG heatmap") is introduced for rapid evaluation and annotation of BCG signal quality.
  • A random forest model is developed to automatically classify high-quality and low-quality BCG data.
• Implementation Steps and Techniques:
  1. Signal Processing and Preprocessing: Noise and motion artifacts are removed using zero-phase filters, and heartbeat intervals (IBI) are preliminarily estimated using a dynamically weighted Bayesian method.
  2. Visualization Annotation Method: A novel "ECG-gated BCG heatmap" is proposed, combining time series and heatmap techniques to annotate BCG signal quality, improving annotation efficiency and consistency.
  3. Machine Learning Classifier Development: Time-domain and frequency-domain features of BCG signals are extracted to train and evaluate a random forest classifier for identifying low-quality data segments.
  4. Quality Filtering and Application: An automatic quality detection algorithm filters low-quality IMU data, significantly improving the accuracy of HR/HRV estimation.

Research Outcomes

• Specific Outcomes:
  • The developed BCG quality classifier achieved an accuracy of 82.56% and an F1 score of 74.17% on an independent test set, effectively distinguishing usable and unusable BCG data segments.
  • After applying quality filtering in non-speech tasks, the error in BCG heartbeat interval estimation was significantly reduced, achieving accuracy comparable to PPG (e.g., mean absolute error decreased from 48.36ms to 39.50ms).
• Comparison with Existing Solutions:
  • PPG signals remain the gold standard for stress monitoring, but this study demonstrates that BCG signals, after quality filtering, are competitive, especially in static and non-speech scenarios.
  • Compared to traditional single-beat quality detection methods, the coarse-grained quality assessment strategy proposed in this study is more cost-effective and efficient.
• Experimental and Evaluation Results:
  • Based on experimental data (81 participants and 8,241 minute-level BCG data segments), the experiments validated that quality filtering significantly improved HR/HRV estimation accuracy and reduced HRV estimation errors in noisy data segments.
  • In specific scenarios (e.g., cold water immersion tasks), the improvements brought by BCG signal quality filtering were particularly notable.
• Limitations and Future Directions:
  • The system's performance in real-world scenarios (e.g., dynamic work environments) has not been fully validated, especially under complex conditions such as frequent motion artifacts and prolonged earbud usage.
  • Further optimization is needed to support different hardware platforms and verify the algorithm's generalizability across various earbud devices.
  • Future research could explore combining IMU and PPG data to enhance the stability of stress monitoring in all-day scenarios.

Conclusion

This study demonstrates the potential of IMU sensors in earbuds for integrated stress assessment solutions, particularly in static and non-speech scenarios, by improving physiological monitoring accuracy through high-quality BCG signals. The developed quality filtering method and visualization technique provide a technological foundation for future applications of earbud devices in closed-loop stress management. Additionally, the study highlights future research directions, such as multi-device collaboration, validation in dynamic scenarios, and application designs that integrate active and passive user interactions.