Estimation of the apnea-hypopnea index in a heterogeneous sleep-disordered population using optimised cardiovascular features

Abstract

Obstructive sleep apnea (OSA) is a highly prevalent sleep disorder, which results in daytime symptoms, a reduced quality of life as well as long-term negative health consequences. OSA diagnosis and severity rating is typically based on the apnea-hypopnea index (AHI) retrieved from overnight poly(somno)graphy. However, polysomnography is costly, obtrusive and not suitable for long-term recordings. Here, we present a method for unobtrusive estimation of the AHI using ECG-based features to detect OSA-related events. Moreover, adding ECG-based sleep/wake scoring yields a fully automatic method for AHI-estimation. Importantly, our algorithm was developed and validated on a combination of clinical datasets, including datasets selectively including OSA-pathology but also a heterogeneous, “real-world” clinical sleep disordered population (262 participants in the validation set). The algorithm provides a good representation of the current gold standard AHI (0.72 correlation, estimation error of 0.56 ± 14.74 events/h), and can also be employed as a screening tool for a large range of OSA severities (ROC AUC ≥ 0.86, Cohen’s kappa ≥ 0.53 and precision ≥70%). The method compares favourably to other OSA monitoring strategies, showing the feasibility of cardiovascular-based surrogates for sleep monitoring to evolve into clinically usable tools.

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

“…The sinus rhythm pulses of good quality were then used to derive the inter-beat intervals (IBIs) necessary for the HRV analysis. Besides the rejection of pulses based on the pulse quality index, we removed an IBI and its preceding IBI when their ratio was larger than 1.5, due to the suspicion of being related to ectopic beats 27 . We derived the amplitude of each sinus rhythm pulse to extract a surrogate respiratory activity signal 15 and, finally, from this surrogate determined the length and amplitude of each breath.…”

“…Our deep learning model had the task of classifying 30-s epochs of each overnight recording as influenced by a respiratory event (RE-epochs, positive class) or not (non-RE epochs, negative class). Similar to our previous research, a 30-s epoch was labelled as RE-epoch if it includes at least 10 s of a respiratory event or if the beginning of an epoch is closer than 5 s to the ending of a respiratory event 27 . The model took all 212 features per epoch as inputs for the entire recording night (maximum 1150 epochs) of each participant, and its output was the probability of each epoch of being a positive class (values from 0 to 1).…”

“…The model took all 212 features per epoch as inputs for the entire recording night (maximum 1150 epochs) of each participant, and its output was the probability of each epoch of being a positive class (values from 0 to 1). The probability threshold to label an epoch as positive class was derived based on the AHI estimation performance as explained in our previous ECG-based research 27 .…”

“…The performance of cardiovascular-based OSA monitoring algorithms is influenced by the presence of other sleep disorders and associated events, and by the types of respiratory events 21 , 22 , 27 . Previously, we reported that other sleep disorders can constitute a confounding factor when detecting respiratory events 21 , 27 . For instance, limb movements tend to be mistaken for respiratory events, while hypopneas, the most common respiratory events, are often difficult to detect compared to apneas.…”

“…Here, we propose a new AHI estimation method developed for wrist-worn, green-light rPPG devices, and asses its performance in a clinical population, comprising healthy subjects and patients with a various types and levels of disordered sleep. In addition to the HRV and movement features used in our previous ECG-based research 27 , we included respiratory activity features, sleep context information in the form of rPPG-based sleep stage probability and feature coverage. A deep learning model employed these features to detect 30-s epochs influenced by respiratory events (RE-epochs), and these RE-epochs were used to estimate the AHI.…”