Wearable monitoring of sleep-disordered breathing: estimation of the apnea–hypopnea index using wrist-worn reflective photoplethysmography

Gabriele B. Papini, Pedro Fonseca, Merel M. van Gilst, J.w.m. Bergmans, Rik Vullings, Sebastiaan OvereemScientific 2020 · DOI

TL;DR

This paper introduces a novel method for estimating the Apnea-Hypopnea Index (AHI), a key metric for diagnosing sleep apnea, using only data from wrist-worn reflective photoplethysmography (rPPG) sensors. Their deep learning approach shows promising results for unobtrusive, long-term monitoring and screening of sleep-disordered breathing, potentially revolutionizing how we diagnose and manage this common condition.

What Problem Does This Paper Address?

Obstructive Sleep Apnea (OSA) is a widespread disorder linked to serious cardiovascular issues. The current gold standard for diagnosis is polysomnography (PSG), a time-consuming, expensive, and intrusive sleep study. This makes it impractical for widespread screening or continuous, long-term monitoring of OSA, especially for tracking its variability over time or in response to treatment. While home sleep apnea tests (HSAT) are less intrusive, they still have limitations. The paper aims to bridge this gap by developing an automatic, unobtrusive method for AHI estimation using readily available wearable technology.

Methods

The core of this research is a deep learning model designed to detect "respiratory event epochs" (RE-epochs) from rPPG signals. These RE-epochs are 30-second segments of sleep identified as potentially containing a respiratory event. The model leverages a rich set of features extracted from the rPPG signal, including:

  • Cardiorespiratory features: Derived from heart rate variability (HRV) and surrogate respiratory activity signals.
  • Movement features: From the accelerometer.
  • Contextual features: Including sleep stage probabilities and feature coverage (indicating data quality).

The model was trained on a large dataset and then evaluated on an independent hold-out set. A key aspect of their methodology is the investigation into the impact of rPPG signal quality, leading to an analysis of results both with and without excluding recordings deemed to have low rPPG quality. This allowed them to assess the robustness of their method and the benefits of data quality control.

Datasets

The study utilized two datasets:

  • SOMNIA dataset: A large collection of simultaneous PSG and unobtrusive sensor recordings from a heterogeneously sleep-disordered population.
  • HealthBed dataset: A smaller set of recordings from healthy adults, used to supplement the training data.

The combined dataset comprised 502 participants. For training and validation, 250 participants were used, with the remaining 252 forming the hold-out set. A subset of 188 participants from the hold-out set was specifically analyzed after excluding recordings with low rPPG quality, to evaluate the impact of data quality.

Key Insights and Ideas

  • rPPG as a viable source for AHI estimation: The study demonstrates that wrist-worn rPPG, a technology commonly found in smartwatches and fitness trackers, can capture sufficient physiological information to estimate AHI, a metric traditionally derived from more complex PSG.
  • Deep learning's power in feature integration: The deep learning model effectively integrates diverse features (cardiorespiratory, movement, sleep stage) to improve RE-epoch detection and subsequent AHI estimation, outperforming simpler approaches.
  • The importance of rPPG signal quality: While the method shows promise even with lower quality data, excluding recordings with poor rPPG quality significantly improves AHI estimation performance, highlighting the need for robust signal processing and quality control in wearable applications.
  • Underestimation vs. Overestimation trade-offs: The method tends to slightly underestimate AHI, which is less clinically critical for screening than overestimation. However, the high ROC AUC values suggest that thresholds can be adjusted to balance sensitivity and specificity for different clinical needs.

Key Results

  • AHI Estimation:
    • With the full hold-out set (252 participants), the estimated AHI correlated with the reference AHI with a Spearman's correlation of 0.61 (p < 0.01) and an estimation error of 3±10 events/h.
  • After excluding low-quality rPPG recordings (188 participants), the correlation improved to 0.67 (p < 0.01).
  • OSA Severity Assessment:
    • The estimated AHI reliably assessed OSA severity, achieving a weighted Cohen's kappa of 0.51.
  • OSA Screening:
    • The method demonstrated good screening performance, with ROC AUC values of 0.84, 0.86, and 0.85 for mild, moderate, and severe OSA, respectively, after excluding low-quality recordings.
  • RE-epoch Detection:
    • The model achieved an accuracy of 86% and a specificity of 94% for RE-epoch detection on the hold-out set. Sensitivity was lower at 39%, particularly for hypopneas (34%).

Limitations and Future Work

  • rPPG Signal Quality: The study acknowledges that rPPG signals can be prone to artefacts from movement and poor skin contact, which can affect feature extraction and model performance. While excluding low-quality recordings improved results, it also reduced the number of usable participants. Future work could focus on developing more robust features that are less susceptible to these artefacts.
  • Underestimation Tendency: The method shows a tendency to underestimate AHI, which could lead to missed diagnoses in some cases. The authors suggest that this is less critical for screening than overestimation and can be managed by adjusting thresholds. However, further investigation into the factors contributing to underestimation, especially in severe OSA cases, is warranted.
  • Generalizability: While the model was tested on a heterogeneous population, further validation on even larger and more diverse datasets, including different age groups and ethnicities, would be beneficial.
  • Comparison to Other Modalities: The authors suggest exploring the applicability of their method to other sensing modalities and investigating how to integrate data from newer wearable sensors (e.g., those measuring oxygen saturation).
  • Clinical Implementation: Future work could focus on refining the model for real-world deployment, potentially by incorporating more advanced signal processing techniques and exploring strategies for continuous, multi-night monitoring.

Summary updated

Abstract

A large part of the worldwide population suffers from obstructive sleep apnea (OSA), a disorder impairing the restorative function of sleep and constituting a risk factor for several cardiovascular pathologies. The standard diagnostic metric to define OSA is the apnea-hypopnea index (AHI), typically obtained by manually annotating polysomnographic recordings. However, this clinical procedure cannot be employed for screening and for long-term monitoring of OSA due to its obtrusiveness and cost. Here, we propose an automatic unobtrusive AHI estimation method fully based on wrist-worn reflective photoplethysmography (rPPG), employing a deep learning model exploiting cardiorespiratory and sleep information extracted from the rPPG signal trained with 250 recordings. We tested our method with an independent set of 188 heterogeneously disordered clinical recordings and we found it estimates the AHI with a good agreement to the gold standard polysomnography reference (correlation = 0.61, estimation error = 3±10 events/h). The estimated AHI was shown to reliably assess OSA severity (weighted Cohen's kappa = 0.51) and screen for OSA (ROC-AUC = 0.84/0.86/0.85 for mild/moderate/severe OSA). These findings suggest that wrist-worn rPPG measurements that can be implemented in wearables such as smartwatches, have the potential to complement standard OSA diagnostic techniques by allowing unobtrusive sleep and respiratory monitoring.
BibTeX
@inproceedings{papini2020wearable,
  author = {Gabriele B. Papini and Pedro Fonseca and Merel M. van Gilst and J.W.M. Bergmans and Rik Vullings and Sebastiaan Overeem},
  title = {{Wearable monitoring of sleep-disordered breathing: estimation of the apnea–hypopnea index using wrist-worn reflective photoplethysmography}},
  booktitle = {Scientific},
  year = {2020},
  doi = {10.1038/s41598-020-69935-7},
}
Figure