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

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.

Figure 1

The selected model architecture for the RE-epoch detection. The numbers below boxes indicates the dimensions (with 1150 being the maximum number of epochs). The rate indicates the drop out rate, N indicates the number of stacked convolution, F the number of filter, K the kernel size, K* the kernel size with dilation rate of 2, C the number of units of the dense layers and std the standard deviation of the Gaussian noise. The block types are further described in the Supplementary section Deep learning model.

Figure 2

Analysis of the estimated AHI performance after removal of low-quality rPPG recordings. (a) Correlation between reference AHI versus estimated AHI; dashed lines delimit the canonical OSA severity classes and the dash-dotted line is the identity line. (b) Bland–Altman plot of the reference AHI and estimated AHI. The bias and the limits of agreement (i.e. 1.96 times the standard deviation of the difference) are shown as events/h. The red and the green dashed lines represent, respectively, the boundaries to define considerable under- and overestimations.

Figure 3

Receiver operating characteristics and confusion matrix of the estimated for the three canonical AHI thresholds after removal of low-quality rPPG recordings. (a) AUC area under each curve; square markers indicate the points in the curve where the estimated AHI threshold for severity classification is equal to the canonical 5, 15 and 30 events/h. (b) OSA severity classes obtained from the AHI (reference severity) and estimated AHI (predicted severity) using the canonical thresholds. In each cell, the percentage per severity is shown (also visually indicated by the color scale) as well as the number of participants.

Figure 4

Characteristics of the considerable underestimated participants that might have influenced the underestimation (for all the participants and for those with at least two class difference between reference and estimated OSA severity). Cardiac comorbidities include bundle-branch-block, premature ventricular/atrial contraction and paroxysmal atrial fibrillation. Cardiovascular medications include anti-arrhythmic compound, ACE-inhibiters, beta-blockers and thyroid hormones.