Direct application of an ECG-based sleep staging algorithm on reflective photoplethysmography data decreases
M M van Gilst1,2, B M Wulterkens3,4, P Fonseca3,4
1Department of Electrical Engineering, Eindhoven University of Technology, PO Box 513, 5600 MB, Eindhoven, The Netherlands. m.m.v.gilst@tue.nl.
BMC Research Notes
|November 10, 2020
Summary
Directly applying sleep staging algorithms trained on ECG data to wrist-worn PPG data reduced performance. Re-training or validating algorithms for each data source is crucial for accurate sleep monitoring.
Area of Science:
- Biomedical Engineering
- Sleep Medicine
- Artificial Intelligence in Healthcare
Background:
- Advancements in neural networks and large sleep datasets drive interest in alternative sleep monitoring.
- Heart rate variability (HRV) from inter-beat intervals (IBIs) is promising for unobtrusive sleep staging.
- Photoplethysmography (PPG) offers a practical alternative for obtaining IBIs, but validation is limited.
Purpose of the Study:
- To evaluate the performance of an ECG-validated sleep staging algorithm when applied directly to PPG-derived IBIs.
- To compare the accuracy of sleep staging using PPG-based HRV versus ECG-based HRV.
Main Methods:
- An automatic sleep staging algorithm, trained and validated on ECG data, was applied to IBIs derived from wrist-worn PPG sensors.
- The algorithm's performance was assessed on 389 polysomnographic recordings from patients with various sleep disorders.
- Performance metrics included kappa statistic and accuracy, compared between PPG and ECG data.
Main Results:
- The algorithm showed moderate agreement with polysomnography but performed significantly lower on PPG-derived HRV compared to ECG-derived HRV (kappa 0.56 vs. 0.60; accuracy 73.0% vs. 75.9%).
- Direct application of algorithms across different data sources can negatively impact performance.
- A statistically significant decrease in performance was observed when using PPG data (p < 0.001).
Conclusions:
- Sleep staging algorithms require validation with each specific data source (e.g., PPG, ECG).
- Re-training algorithms for PPG-derived data is recommended to maintain performance.
- This highlights the need for source-specific algorithm validation in sleep monitoring.
