AIOSA: An approach to the automatic identification of obstructive sleep apnea events based on deep learning
Introduction
Among sleep-related breathing disorders, Obstructive Sleep Apnea Syndrome (OSAS) is the most common one [1]. It is caused by an increased upper airway resistance during sleep, leading to episodes of partial or complete interruption of airflow, that bring to phasic reductions in blood oxygen content; arousals from sleep are usually required to interrupt these events. OSAS commonly manifests itself with excessive daytime sleepiness due to sleep fragmentation; however, its most relevant health-related burden is represented by an increased risk of cardiovascular and cerebrovascular accidents such as myocardial infarction and ischemic stroke [2].
The gold standard test for diagnosing OSAS is polysomnography (PSG), which requires overnight recording of at least the following parameters: airflow, blood oxygen saturation, thoracic and abdominal movements. Moreover, some or all of the following additional parameters are often recorded: snoring, electrocardiography, electroencephalography, electrooculography, surface electromyography of the mylohyoid and tibialis anterior muscles. Such recordings are then manually tagged by a trained physician against the presence of apneic events (Fig. 1). As a result, performing a PSG is labour-, time-, and money-consuming. Respiratory events can be classified as apneas and hypopneas based on PSG features: the former are characterized by a >90% reduction of respiratory flow for at least 10 s, whereas the latter require a >30% reduction of respiratory flow for at least 10 s with a concomitant reduction in blood oxygen saturation ≥3% [3]. The severity of OSAS is graded by means of a composite measure, named apnea-hypopnea index (AHI), which is calculated dividing the sum of all apneas and hypopneas by the total hours of sleep. OSAS is defined as mild when 5≤AHI<15, moderate when 15≤AHI<30, and severe when AHI≥30 [4].
The detection and treatment of OSAS are particularly important in patients who suffered a stroke [5]. Stroke is defined as an episode of neurologic dysfunction due to infarction (ischemic stroke) or focal collection of blood (hemorrhagic stroke) within the central nervous system [6], and represents the second cause of death and the third cause of disability worldwide [7]. The optimal inpatient setting for acute stroke patients is represented by specialized semi-intensive care wards, named stroke units [8]. In a stroke unit, all patients undergo continuous monitoring of many vital parameters such as noninvasive blood pressure, multi‑lead electrocardiography, photoplethysmography-derived blood oxygen saturimetry, and thoracic impedance-derived respiratory rate.
The prevalence of OSAS is high in the general population, with 49.7% of men and 23.4% of women suffering from moderate to severe OSAS [9]. In patients with acute stroke, OSAS is even more prevalent, with up to 91.2% of patients being affected and 44.6% experiencing severe OSAS [10]. After an acute stroke, the presence of severe OSAS is associated with higher mortality, worse neurological deficits, worse functional outcome after rehabilitation, and a higher likelihood of uncontrolled hypertension [11,12].
The cornerstone of OSAS treatment is represented by nocturnal continuous positive airway pressure (CPAP) ventilation [13]. This noninvasive ventilation system increases air pressure in the upper respiratory tract, thus preventing airway collapse. Treating patients with CPAP determines an improved functional outcome and a reduced 5-year mortality risk due to cardiovascular disease [5].
Unfortunately, performing a PSG in an electrically hostile environment, such as a stroke unit, on neurologically impaired patients is a difficult task, with the result that signals are often affected by noise (Fig. 2); moreover, the number of strokes per day vastly outnumbers the availability of polysomnographs and dedicated healthcare professionals. Therefore, a simple and automated recognition system to identify OSAS cases among acute stroke patients is highly desirable.
The continuous multiparametric recording of vital signs that is routinely performed in stroke units represents a relevant data source for a comprehensive assessment of patients' health status. However, as stated in [3], diagnosing OSAS with traditional manual sleep scoring requires explicit evaluation of parameters not recorded during stroke unit monitoring, like, for instance, airflow and thoracoabdominal movements.
Automated analysis of the simplified stroke unit monitoring system may reveal implicit features and thus allow reliable OSAS screening, with no additional procedures or sensors being required and at no extra cost. Similar approaches have been tried before (see Section 2), with variable success. However, those experiments were performed on data recorded in ideal conditions and on highly selected patients, with stringent exclusion criteria regarding cardiac, respiratory, and other comorbidities. Results obtained under such experimental conditions are hardly generalizable to real-life circumstances, where this solution would be of actual use. In addition, typical approaches are only able to establish whether a patient is affected by OSAS or to roughly locate the presence of anomalous respiratory events during sleep following a coarse-grained windowing or segmentation strategy.
In this paper, we develop a convolutional-based deep learning framework that deals with waveform data by effectively summarizing them and extracting their key properties. Such an architecture can be used as a component in larger models to preprocess raw signals before further elaboration. Unlike previous deep learning solutions applied to OSAS detection, the proposed architecture is specifically designed to handle and summarize raw signals with an arbitrarily high sampling frequency, preserving temporal relationships over long time windows. Moreover, to the best of our knowledge, for the first time apnea events are tagged at one-second granularity. Such an ability provides physicians with fine-grained information about the condition of the patient, allowing them to better interpret and validate the results of the model.
We apply the proposed framework to the well-known Apnea-ECG Database [14], outperforming current state-of-the-art solutions. Then, we turn to a real case scenario, considering the task of detecting OSAS events during sleep in a stroke unit, with the goal of identifying serious cases. Unlike what happens with existing solutions, the data is collected from the monitoring of unselected patients and include electrocardiogram (ECG) and peripheral oxygen saturation (SpO2). The system is intended to work in an offline fashion, processing overnight recordings as a whole, as typically done in the field. The achieved results are deemed to be satisfactory by domain experts, and may be interpreted as an indicator of the general applicability of the approach in a production setting. This is particularly meaningful as a trained physician necessarily has to rely on the more complex polysomnograph data to perform a similar OSAS assessment. The choice of relying on deep learning instead of classical machine learning techniques is motivated by the fact that, as witnessed in the literature [15], deep learning models perform automatic feature extraction. This is of great help since, based on a series of meetings with expert physicians, it emerged that identifying a set of hand-engineered attributes from raw data is quite challenging.
The main contributions of the work are the following:
- •
the design of a novel neural network architecture able to assess OSAS severity and tag apnea events at one-second granularity;
- •
the network ability to summarize raw physiological signals, reducing their temporal resolution while effectively preserving temporal relationships over long time windows, thanks to the usage of dilated convolutions arranged in a pondered pyramidal scheme;
- •
the validation of the proposed model on a well-established testbed, that confirms its superiority with respect to existing solutions;
- •
the experimental evaluation of the model on a novel Stroke Unit dataset, consisting of data pertaining to unselected patients affected by multiple comorbidities, which suggests the effectiveness of our solution in terms of both classification performance and clinical interpretability of the model output.
The paper is organized as follows. Section 2 presents the state of the art in the automatic detection of sleep apnea events. Section 3 illustrates the considered domains. In addition, it describes the architectures of the models and the design of the experiments. Section 4 reports the results obtained from both the Apnea-ECG Database and our dataset. Conclusions provide an assessment of the work done, and outline future research directions.
Section snippets
Related work
A large number of approaches to the automatic identification of sleep apnea and hypopnea events have been proposed in the literature.
In Table 1, Table 2, we provide a concise, but comprehensive, account of the approaches that make use of ECG and SpO2 recordings, being the far more used and those that we consider in our work. As it is well known in the medical domain [16,17], information conveyed by such signals is strongly related to the presence of OSAS.
For each entry, we report information
Materials and methods
In this section, we describe the datasets, the developed models, the experimental setting, and the proposed neural network architecture. Overall, our work differs from existing ones in at least three fundamental aspects: (i) we focus on a real-world scenario, where patient exclusion criteria are much less stringent than those usually applied; (ii) we tag apnea and hypopnea events at a 1-second granularity, considerably enhancing the interpretability of the output; and, (iii) a distinctive
Results
In this section, the outcomes of the evaluation of the proposed models on both the Apnea-ECG database and our Stroke Unit dataset are reported.
Conclusions and future work
The ultimate goal of the present work was to develop an effective and easily deployable tool to help the clinical decision-making process in the context of OSAS. To this end, we proposed a deep learning framework for the detection of sleep apnea events, based on convolutional neural networks. Its distinctive feature is that it is able to deal with waveform data, such as physiological signals, by effectively summarizing them and extracting their key components. We first applied the framework to
Declaration of competing interest
The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Personnel belonging to the University of Udine and to the Udine University Hospital, Italy.
Acknowledgements
This work was supported by Google Academic Research Grant and TensorFlow Research Cloud programs, as well as the Italian INdAM-GNCS project Ragionamento Strategico e Sintesi Automatica di Sistemi Multi-Agente. Moreover, the authors would like to thank the reviewers for their valuable comments and suggestions, which helped them in improving the paper.
References (64)
- et al.
Prevalence of obstructive sleep apnea in the general population: a systematic review
Sleep Med Rev
(2017) - et al.
Prevalence of sleep-disordered breathing in the general population: the HypnoLaus study
Lancet Respir Med
(2015) - et al.
Prevalence of sleep apnea at the acute phase of ischemic stroke with or without thrombolysis
Sleep Med
(2017) - et al.
Study of association of severity of sleep disordered breathing and functional outcome in stroke patients
Sleep Med
(2017) - et al.
Detection of abnormal respiratory events with single channel ECG and hybrid machine learning model in patients with obstructive sleep apnea
IRBM
(2020) - et al.
An algorithm for sleep apnea detection from single-lead ECG using hermite basis functions
Comput Biol Med
(2016) - et al.
A method to detect sleep apnea based on deep neural network and hidden markov model using single-lead ECG signal
Neurocomputing
(2018) - et al.
Obstructive sleep apnea screening from unprocessed ECG signals using statistical modelling
Biomed Signal Process Control
(2021) - et al.
Heart rate variability feature selection in the presence of sleep apnea: an expert system for the characterization and detection of the disorder
Comput Biol Med
(2017) - et al.
Feature selection from nocturnal oximetry using genetic algorithms to assist in obstructive sleep apnoea diagnosis
Med Eng Phys
(2012)
Classification techniques on computerized systems to predict and/or to detect apnea: a systematic review
Comput Methods Programs Biomed
An introduction to ROC analysis
Pattern Recogn Lett
Obstructive sleep apnoea and cardiovascular disease
Lancet Respir Med
Rules for scoring respiratory events in sleep: update of the 2007 AASM manual for the scoring of sleep and associated events. Deliberations of the sleep apnea definitions task force of the American Academy of Sleep Medicine
J Clin Sleep Med
Screening for obstructive sleep apnea in adults: an evidence review for the U.S. preventive services task Force, U.S. preventive services task force evidence syntheses, formerly systematic evidence reviews
CPAP as treatment of sleep apnea after stroke: a meta-analysis of randomized trials
Neurology
An updated definition of stroke for the 21st century: a statement for healthcare professionals from the American Heart Association/American Stroke Association
Stroke
Global burden of disease
Guidelines for the early management of patients with acute ischemic stroke: a guideline for healthcare professionals from the American Heart Association/American Stroke Association
Stroke
Obstructive sleep apnea and serious adverse outcomes in patients with cardiovascular or cerebrovascular disease: a PRISMA-compliant systematic review and meta-analysis
Medicine (Baltimore)
Continuous positive airways pressure for obstructive sleep apnoea in adults
Cochrane Database Syst Rev
The apnea-ECG database
Deep learning
The correlation between oxygen saturation indices and the standard obstructive sleep apnea severity
Ann Thorac Med
Heart rate variability and obstructive sleep apnea: current perspectives and novel technologies
J Sleep Res
Automated scoring of obstructive sleep apnea and hypopnea events using short-term electrocardiogram recordings
IEEE Trans Inf Technol Biomed
Automated detection of obstructive sleep apnea events from a single-lead electrocardiogram using a convolutional neural network
J Med Syst
Deep learning approaches for automatic detection of sleep apnea events from an electrocardiogram
Comput Methods Prog Biomed
Obstructive sleep apnea prediction from electrocardiogram scalograms and spectrograms using convolutional neural networks
Physiol Meas
Automated detection of obstructive sleep apnoea by single-lead ECG through ELM classification
An obstructive sleep apnea detection approach using a discriminative hidden markov model from ECG signals
IEEE Trans Biomed Eng
Recurrent neural network based classification of ECG signal features for obstruction of sleep apnea detection
Cited by (24)
Accurate apnea and hypopnea localization in PSG with Multi-scale object detection via Dual-modal Feature Learning
2024, Biomedical Signal Processing and ControlApneaNet: A hybrid 1DCNN-LSTM architecture for detection of Obstructive Sleep Apnea using digitized ECG signals
2023, Biomedical Signal Processing and ControlA review of automated sleep disorder detection
2022, Computers in Biology and MedicineDeep Learning Approaches for Early Detection of Obstructive Sleep Apnea Using Single-Channel ECG: A Systematic Literature Review
2024, Communications in Computer and Information Science
- 1
Main authors.