TL;DR
This study explores how physiological signals from wrist-worn wearables can differentiate pregnant from non-pregnant individuals. It finds that pulse wave morphology features, particularly systolic pulse wave deterioration, are more effective than traditional heart rate variability (HRV) metrics for this distinction. Analyzing data from deep sleep periods further enhances this differentiation, suggesting a promising avenue for non-invasive maternal health monitoring.
What Problem Does This Paper Address?
Pregnancy involves significant physiological changes, and disruptions in the maternal autonomic nervous system (ANS) can signal potential complications like hypertensive disorders or gestational diabetes. Early detection of these issues is crucial for better maternal and perinatal outcomes. While wearable devices can non-invasively track autonomic regulation through photoplethysmography (PPG), it's unclear which specific features derived from PPG are most valuable for monitoring maternal health. This paper aims to identify which heart rate variability (HRV) and pulse wave morphology features best distinguish healthy pregnant women from non-pregnant women, hypothesizing that these features reflect physiological adaptations to pregnancy and could be altered in complicated pregnancies.
Methods
The researchers employed a binary classification approach using logistic regression and stepwise forward feature elimination. They extracted a comprehensive set of both HRV and morphological features from nighttime PPG recordings. The core idea was to identify features that most effectively discriminate between the two groups (pregnant vs. non-pregnant). This process was repeated for HRV features alone, morphological features alone, and a combination of both. To refine the analysis, they also stratified the data by sleep stages (light sleep, deep sleep, REM) to see if specific sleep phases offered better discriminatory power.
Datasets
The study utilized two retrospective datasets:
- Pregnant group: Data from 45 healthy women in their second or third trimester of pregnancy, collected over two nighttime sessions approximately eight weeks apart.
- Non-pregnant group: Data from healthy women of childbearing age, collected during a sleep study.
For this study, the pregnant group's PPG data was downsampled to 32 Hz to match the non-pregnant group's sampling rate, enabling direct comparison.
Key Insights and Ideas
- Morphology over HRV: Pulse wave morphology features were found to be more valuable for discriminating between pregnant and non-pregnant women than traditional HRV features. This suggests that the physical characteristics of the pulse wave carry significant information about pregnancy-related physiological changes.
- Deep Sleep is Key: Analyzing data specifically from deep sleep (N3) periods significantly enhanced the differences between pregnant and non-pregnant women, outperforming analyses based on the entire night or other sleep stages. This highlights the potential of focusing on specific physiological states for more sensitive monitoring.
- Systolic Pulse Wave Deterioration: The "systolic pulse wave deterioration" emerged as the single most valuable feature for differentiation, followed closely by mean heart rate. This points to specific aspects of the pulse wave's shape being highly indicative of pregnancy status.
- HRV Limitations at Low Sampling Rates: The study notes that the relatively low sampling rate (32 Hz) of PPG data might limit the precision of some HRV features, particularly frequency-domain ones. This suggests that while PPG is convenient, its limitations need to be considered when interpreting HRV.
Key Results
- Overall Discrimination: Morphological features achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.825, while HRV features achieved 0.74.
- Top Features: The most important features for discrimination were identified as Systolic Pulse Duration (SPD) and Mean Heart Rate (HR).
- Sleep Stage Impact: When analyzing data from deep sleep (N3), the classification performance was notably enhanced, suggesting this stage is particularly informative.
- Combined Approach: While morphological features alone performed well, combining them with HRV features generally yielded comparable or slightly improved results compared to morphology alone.
Limitations and Future Work
- Sampling Rate: The study acknowledges the limitation of the 32 Hz PPG sampling rate, which may affect the accuracy of certain HRV features. Future work could explore higher sampling rates or exclude features most susceptible to low-rate errors.
- Age Difference: A median age difference of six years exists between the pregnant and non-pregnant groups. While the study investigated age's impact and found it less influential on key morphological features, a prospective, age-matched study is recommended.
- Wrist Position: The influence of wrist position on PPG amplitude was noted as a potential confounder, though its impact is deemed less significant than pregnancy itself.
- Automated Sleep Scoring: The use of an automated sleep scoring algorithm, rather than technician-based polysomnography, is a limitation, though it offers advantages for free-living monitoring.
- Future Directions: The authors suggest incorporating PRSA and HRF features into future maternal health assessments. They also emphasize the potential of morphological features for identifying pregnancy complications like preeclampsia and recommend focusing on deep sleep data for more sensitive monitoring.
Summary updated