TL;DR
This paper introduces a novel method to estimate SDNN24, a crucial metric for cardiovascular health, using only heart rate (HR) data commonly available from wrist-worn wearables. This approach bypasses the need for precise inter-beat interval (IBI) data, which is often corrupted by motion artifacts in wearable devices, making continuous cardiac health monitoring more feasible.
What Problem Does This Paper Address?
The standard deviation of inter-beat intervals over 24 hours (SDNN24) is a gold standard for assessing cardiovascular health. Traditionally, this metric is derived from electrocardiogram (ECG) data, often collected using Holter monitors. However, these devices are cumbersome and not suitable for continuous, everyday monitoring. Modern wrist-worn wearables, like fitness trackers, offer continuous HR monitoring, which is less susceptible to motion artifacts than raw IBI data. The challenge has been to accurately estimate SDNN24 from this readily available HR data, as the traditional methods rely on precise IBI measurements that are difficult to obtain reliably from wearables. This gap prevents widespread use of wearable data for comprehensive cardiac health assessment.
Methods
The core innovation lies in estimating SDNN24 from HR data, which is essentially a low-pass filtered version of the IBI signal. The authors propose several methods:
- Time Domain Analysis: They calculate SDNN24 using various HR-derived metrics, including SDNNi24 (mean of segment SDNNs), SDANN24 (SD of segment means), and importantly, SDANNHR24 (SD of average NN intervals derived from HR).
- Frequency Domain Analysis: They leverage the fact that SDNN24 represents the total power in the heart rate variability (HRV) spectrum. Since HR data acts as a low-pass filter, it primarily captures the lower frequency components (Ultra Low Frequency - ULF and Very Low Frequency - VLF), which are known to be strongly influenced by the sinoatrial node and circadian rhythms. To compensate for the missing high frequencies, they develop adjusted methods (SDANNHR24adjMean and SDANNHR24adjw) that incorporate estimated power from higher frequencies.
- Machine Learning: They employ Logistic Regression (LR) and Neural Networks (NN) using HR-derived features (including circadian rhythm parameters like MESOR and Amplitude) to predict SDNN24 and discriminate between healthy and unhealthy subjects.
Datasets
The study utilized three publicly available PhysioNet datasets:
- nsr2db: Normal Sinus Rhythm RR Interval Database (54 subjects, ~23 hours of ECG).
- chfdb: Congestive Heart Failure RR Interval Database (29 subjects with heart failure, ~23 hours of ECG).
- mmash: Multilevel Monitoring of Activity and Sleep in Healthy people (22 healthy subjects, 24 hours of beat-to-beat heart data, accelerometer, sleep quality, etc.).
The MMASH dataset was particularly important for simulating HR data from wearables and estimating the error introduced by PPG sensors.
Key Insights and Ideas
- HR as a Proxy for Low-Frequency HRV: The study highlights that HR, being a low-pass filtered version of IBI, is a stable proxy for the crucial low-frequency components (ULF and VLF) of HRV. These low frequencies are strongly linked to the sinoatrial node's activity and circadian rhythms, which are key indicators of cardiac health.
- SDNN24 Underestimation from HR: Directly calculating SDNN24 from HR data (e.g., SDANNHR24) inherently underestimates the true value because high-frequency components are lost. However, the paper demonstrates that this underestimation is predictable and can be corrected.
- Low Frequencies Dominate Cardiovascular Risk: The power spectrum analysis revealed significant differences in ULF and VLF between low and high cardiovascular risk groups, while LF and HF showed similar patterns. This suggests that low-frequency HRV components are more discriminative for assessing cardiovascular risk.
- HR Features for Risk Stratification: Machine learning models trained on HR-derived features (especially circadian rhythm parameters) showed high accuracy in discriminating between healthy and unhealthy individuals, often outperforming models based solely on SDNN24.
Key Results
- Accurate SDNN24 Estimation: The adjusted methods (SDANNHR24adjMean and SDANNHR24adjw) were able to accurately estimate SDNN24, with SDANNHR24adjw showing no statistical difference from the gold standard SDNN24 across various segment lengths.
- Discrimination of Cardiovascular Risk: SDNNHR24 (derived from HR) was found to be slightly more informative than SDNN24 for discriminating between healthy and unhealthy subjects, particularly when using shorter time segments.
- Machine Learning Performance: Logistic Regression models using HR features (LRHR) achieved high precision (0.94), recall (1.00), and F1-score (0.97) for predicting low-risk individuals, and comparable performance for high-risk individuals. Fully connected neural networks also performed well.
- Error Reduction with Longer Time Windows: For HR estimation error, longer time windows (e.g., 60 minutes) resulted in lower bias variability, indicating that averaging HR over longer periods smooths out motion artifact-induced errors.
Limitations and Future Work
- Simulated HR Data: A primary limitation is that the HR data used for validation was largely simulated from ECG data, with added error to mimic wearable devices. While this approach allowed for controlled experimentation, the authors acknowledge that future work needs to validate their methods on real-world data from wrist-worn devices.
- Generalizability: While the study used diverse datasets, further validation across a wider range of populations and device types is recommended.
- Interpretability vs. Performance: The authors suggest that while neural networks achieve high predictive power, simpler models like Logistic Regression (LRHR) are preferable due to their interpretability, especially when performance is comparable.
The study strongly suggests that continuous HR monitoring from wearables is sufficient for robust SDNN24 estimation and cardiovascular risk assessment, paving the way for more accessible and proactive cardiac health management.
Summary updated