Authors
Larry Zhang, Viswam Nathan, Cristina Rosa, Jilong Kuang, Wendy Berry Mendes, Jun Alex Gao
Abstract
Stress monitoring has become a focal interest in health sensing due to the mental and physical effects of long-term stress. Recent work demonstrated the feasibility of photoplethysmography (PPG)-based heart rate variability (HRV) features from earbud sensors to detect stress. However, morphological PPG features from earbuds have not been evaluated for stress detection. We analyzed physiological data from periods of stress and non-stress for 97 subjects. We trained machine learning models on PPG morphological features and HRV features from the earbuds, as well as ECG HRV features from a reference device. The morphological features (F1 score: 0.879) outperformed PPG HRV features (F1 score: 0.773) in stress classification. The combination of PPG morphological features and HRV features (F1 score: 0.880) performed similarly to ECG HRV features (F1 score: 0.887; ΔF1% = 0.798%). The results suggest earbud PPG morphological and HRV features can detect stress with similar fidelity to ECG, despite the smaller form factor and limited sampling rate. Thus, earbud sensors may be a strong candidate for stress monitoring in physiology due to their user-friendly and comfortable nature.
Key Findings
- Morphological features outperformed HRV features: PPG morphological features achieved an F1 score of 0.879 vs. 0.773 for HRV features
- Quality checking improved performance: Signal quality checks increased morphological feature performance by 17.92% (from 0.750 to 0.870 F1 score)
- Performance comparable to ECG: Combined morphological and HRV features achieved 0.880 F1 score, within 0.798% of reference ECG performance (0.887 F1 score)
- Timing features most important: Pulse Time, MS Time, and Crest Time had higher feature importance than amplitude features
- Earbud sensors are viable: Despite 25 Hz sampling rate and small form factor, earbud sensors achieved comparable performance to ECG monitoring
- Individual reactivity captured: Morphological features captured individual differences in stress reactivity from baseline
Methodology
The study used 97 healthy participants (ages 18-51) with validated stress protocols:
- Data Collection: 5-minute baseline periods, stress tasks (mental arithmetic, public speaking), music recovery periods, and questionnaires
- Signal Processing: PPG signals bandpass filtered (0.7-3.5 Hz) and segmented into 1-minute chunks
- Feature Extraction:
- Morphological: Pulse Time, Crest Time, MS Time, Pulse Wave Amplitude, MS Amplitude
- HRV: Heart Rate, RMSSD, SDNN, RSA
- Quality Control: Waveform quality checks based on peak detection to exclude poor-quality signals
- Machine Learning: Random Forest, Logistic Regression, and SVM models with leave-one-subject cross-validation
- Reference: Validated against manually-scored ECG data from Biopac equipment (1000 Hz sampling)
Impact
This research demonstrates morphological PPG analysis for wearable stress monitoring:
- Wearable technology: Shows that earbud sensors can achieve performance comparable to clinical-grade devices
- Continuous monitoring: Enables stress monitoring in natural settings for preventive healthcare
- Clinical applications: Provides foundation for stress management tools and cardiovascular disease prevention
- Consumer integration: Supports integration of stress monitoring into consumer devices like wireless earbuds
- Signal processing: The quality checking methodology improves reliability of morphological feature extraction
- Research framework: Establishes protocols for evaluating PPG morphology in stress detection
- Healthcare access: Offers an alternative to ECG-based stress monitoring