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

Gabriele B. Papini, Pedro Fonseca, Merel M. van Gilst, Johannes van Dijk, Dirk Pevernagie, J.w.m. Bergmans, Rik Vullings, Sebastiaan OvereemScientific 2019 · DOI

TL;DR

This paper introduces a novel, fully automatic method for estimating the Apnea-Hypopnea Index (AHI), a key metric for diagnosing and assessing the severity of Obstructive Sleep Apnea (OSA), using only electrocardiogram (ECG) data. By leveraging optimized cardiovascular features and an ECG-based sleep/wake scoring system, the method offers a promising unobtrusive alternative to traditional, costly polysomnography, demonstrating good accuracy and screening capabilities across a diverse patient population.

What Problem Does This Paper Address?

Obstructive Sleep Apnea (OSA) is a widespread sleep disorder with significant negative impacts on quality of life and long-term health. The standard diagnostic tool, polysomnography (PSG), is expensive, cumbersome, and not practical for long-term monitoring or widespread screening. While home sleep tests (HST) and questionnaires exist, they have limitations in cost, obtrusiveness, or accuracy, particularly for individuals with subtle symptoms or diverse sleep disorders. This leaves a significant portion of the population undiagnosed or undertreated. The paper aims to bridge this gap by developing an accurate, unobtrusive, and automated method for AHI estimation using readily available ECG signals.

Methods

The core of the proposed method lies in an epoch-by-epoch logistic regression classifier that estimates the probability of a 30-second epoch containing a respiratory event (RE-epoch). This classifier is trained on a rich set of ECG-derived features, including heart rate variability (HRV) metrics and activity counts. A key innovation is the feature optimization framework, which employs a greedy search to apply a series of transformations to the raw features, aiming to maximize their discriminative power for RE-epochs.

To achieve a fully automatic system, the researchers integrated an ECG-based sleep/wake scoring algorithm. This algorithm, combined with the RE-epoch detection, allows for AHI estimation without manual sleep staging. The AHI is then calculated by correlating the number of detected RE-epochs with the total sleep epochs, with a multiplicative coefficient derived from linear regression to account for the non-one-to-one relationship between RE-epochs and actual respiratory events.

Datasets

The study utilized a diverse collection of five datasets to ensure the robustness and generalizability of their method:

  • UCD dataset: Overnight PSG recordings from adults with suspected sleep-related breathing disorders.
  • Apnea-ECG dataset: A well-known dataset for OSA research, containing 70 overnight recordings.
  • HealthBed dataset: Recordings from healthy individuals without sleep complaints.
  • Auto-PSG dataset: Recordings from a clinical sleep laboratory population, representative of OSA cases.
  • SOMNIA dataset: Recordings from patients with a wide range of sleep disorders, including insomnia, movement disorders, parasomnias, and OSA, reflecting a "real-world" clinical scenario.

The pooled validation set comprised 262 participants. The Auto-PSG, HealthBed, and SOMNIA datasets were used for training and independent validation within a cross-validation scheme, while the UCD and Apnea-ECG datasets were used solely for validation to strengthen performance evaluation.

Key Insights and Ideas

  • ECG as a powerful surrogate: The study strongly validates the potential of ECG signals, beyond just heart rate, to capture complex physiological changes associated with sleep-disordered breathing.
  • Feature optimization is crucial: The "feature optimization framework" that systematically applies transformations to ECG features significantly enhances their ability to detect respiratory events, suggesting that raw features are insufficient on their own.
  • Heterogeneity is key for real-world applicability: Validating the model on a diverse population, including those with various sleep disorders beyond OSA, is critical for developing tools that can be used in general clinical practice.
  • Automatic sleep/wake scoring is feasible and minimally impactful: Integrating an ECG-based sleep/wake classifier, while slightly reducing performance metrics, enables a fully automated system with minimal loss of accuracy, making it more practical.
  • Hypopneas are harder to detect than apneas: The research highlights that hypopneas, characterized by less pronounced cardiovascular responses, are more challenging to detect from ECG alone compared to obstructive apneas.

Key Results

The proposed method demonstrated promising results:

  • AHI Estimation: A good representation of the gold standard AHI was achieved, with a Spearman's correlation coefficient of 0.72 and an estimation error of 0.56 ± 14.74 events/h.
  • OSA Screening: The method performed well as a screening tool for various OSA severities, achieving ROC AUC ≥ 0.86 and Cohen's kappa ≥ 0.53.
  • Performance Comparison: The method favorably compared to other OSA monitoring strategies, including questionnaires like STOP-Bang and Berlin, showing higher accuracy for screening mild and moderate OSA.
  • Fully Automatic System: The fully automatic version (with ECG-based sleep/wake scoring) showed only a small decrease in performance compared to manual scoring, with a correlation of 0.72 and an average bias of 0.56 events/h.
  • Severity Classification: The method was able to assign correct OSA severity classes in most cases, with only a tendency to underestimate mild and moderate cases by one level.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Underestimation in Severe OSA: The method showed a tendency to underestimate AHI in participants with severe OSA, potentially due to the training datasets having fewer extremely severe cases. Future work could involve training with more extreme OSA data.
  • Impact of Comorbidities: Participants with comorbid sleep disorders (e.g., insomnia, movement disorders) sometimes led to overestimation of AHI. This suggests that while the method is robust, further refinement might be needed for complex cases.
  • Hypopnea Detection Challenges: The difficulty in detecting hypopneas remains a challenge, possibly due to their subtler cardiovascular impact. Future research could explore more advanced signal processing or feature extraction for hypopneas.
  • Irregular Heartbeats: The method's reliance on HRV features makes it less suitable for individuals with a high prevalence of irregular heartbeats. Future work could focus on developing methods to handle such cases or using them as indicators for further investigation.
  • Sensor Agnosticism: The authors envision translating this method to other sensors like smartwatches and ECG patches, enabling fully automatic, unobtrusive, and long-term OSA monitoring.
  • Beyond AHI: While the current focus is on AHI estimation, future research could explore deriving other relevant metrics for OSA monitoring from the ECG signal.

Summary updated

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.
BibTeX
@inproceedings{papini2019estimation,
  author = {Gabriele B. Papini and Pedro Fonseca and Merel M. van Gilst and Johannes van Dijk and Dirk Pevernagie and J.W.M. Bergmans and Rik Vullings and Sebastiaan Overeem},
  title = {{Estimation of the apnea-hypopnea index in a heterogeneous sleep-disordered population using optimised cardiovascular features}},
  booktitle = {Scientific},
  year = {2019},
  doi = {10.1038/s41598-019-53403-y},
}
Figure