Ensemble-learning regression to estimate sleep apnea severity using at-home oximetry in adults

Deep learning of sleep apnea-hypopnea events for accurate classification of obstructive sleep apnea and determination of clinical severity

BACKGROUND

/Objective: Automatic apnea/hypopnea events classification, crucial for clinical applications, often faces challenges, particularly in hypopnea detection. This study aimed to evaluate the efficiency of a combined approach using nasal respiration flow (RF), peripheral oxygen saturation (SpO2), and ECG signals during polysomnography (PSG) for improved sleep apnea/hypopnea detection and obstructive sleep apnea (OSA) severity screening.

METHODS

An Xception network was trained using main features from RF, SpO2, and ECG signals obtained during PSG. In addition, we incorporated demographic data for enhanced performance. The detection of apnea/hypopnea events was based on RF and SpO2 feature sets, while the screening and severity categorization of OSA utilized predicted apnea/hypopnea events in conjunction with demographic data.

RESULTS

Using RF and SpO2 feature sets, our model achieved an accuracy of 94 % in detecting apnea/hypopnea events. For OSA screening, an exceptional accuracy of 99 % and an AUC of 0.99 were achieved. OSA severity categorization yielded an accuracy of 93 % and an AUC of 0.91, with no misclassification between normal and mild OSA versus moderate and severe OSA. However, classification errors predominantly arose in cases with hypopnea-prevalent participants.

CONCLUSIONS

The proposed method offers a robust automatic detection system for apnea/hypopnea events, requiring fewer sensors than traditional PSG, and demonstrates exceptional performance. Additionally, the classification algorithms for OSA screening and severity categorization exhibit significant discriminatory capacity.

Prediction of the Sleep Apnea Severity Using 2D-Convolutional Neural Networks and Respiratory Effort Signals

The high prevalence of sleep apnea and the limitations of polysomnography have prompted the investigation of strategies aimed at automated diagnosis using a restricted number of physiological measures. This study aimed to demonstrate that thoracic (THO) and abdominal (ABD) movement signals are useful for accurately estimating the severity of sleep apnea, even if central respiratory events are present. Thus, we developed 2D-convolutional neural networks (CNNs) jointly using THO and ABD to automatically estimate sleep apnea severity and evaluate the central event contribution. Our proposal achieved an intraclass correlation coefficient (ICC) = 0.75 and a root mean square error (RMSE) = 10.33 events/h when estimating the apnea-hypopnea index, and ICC = 0.83 and RMSE = 0.95 events/h when estimating the central apnea index. The CNN obtained accuracies of 94.98%, 79.82%, and 81.60% for 5, 15, and 30 events/h when evaluating the complete apnea hypopnea index. The model improved when the nature of the events was central: 98.72% and 99.74% accuracy for 5 and 15 events/h. Hence, the information extracted from these signals using CNNs could be a powerful tool to diagnose sleep apnea, especially in subjects with a high density of central apnea events.

Deep learning for obstructive sleep apnea diagnosis based on single channel oximetry

Obstructive sleep apnea (OSA) is a serious medical condition with a high prevalence, although diagnosis remains a challenge. Existing home sleep tests may provide acceptable diagnosis performance but have shown several limitations. In this retrospective study, we used 12,923 polysomnography recordings from six independent databases to develop and evaluate a deep learning model, called OxiNet, for the estimation of the apnea-hypopnea index from the oximetry signal. We evaluated OxiNet performance across ethnicity, age, sex, and comorbidity. OxiNet missed 0.2% of all test set moderate-to-severe OSA patients against 21% for the best benchmark.

Obstructive sleep apnea event detection using explainable deep learning models for a portable monitor

Background

Polysomnography (PSG) is the gold standard for detecting obstructive sleep apnea (OSA). However, this technique has many disadvantages when using it outside the hospital or for daily use. Portable monitors (PMs) aim to streamline the OSA detection process through deep learning (DL).

Materials and methods

We studied how to detect OSA events and calculate the apnea-hypopnea index (AHI) by using deep learning models that aim to be implemented on PMs. Several deep learning models are presented after being trained on polysomnography data from the National Sleep Research Resource (NSRR) repository. The best hyperparameters for the DL architecture are presented. In addition, emphasis is focused on model explainability techniques, concretely on Gradient-weighted Class Activation Mapping (Grad-CAM).

Results

The results for the best DL model are presented and analyzed. The interpretability of the DL model is also analyzed by studying the regions of the signals that are most relevant for the model to make the decision. The model that yields the best result is a one-dimensional convolutional neural network (1D-CNN) with 84.3% accuracy.

Conclusion

The use of PMs using machine learning techniques for detecting OSA events still has a long way to go. However, our method for developing explainable DL models demonstrates that PMs appear to be a promising alternative to PSG in the future for the detection of obstructive apnea events and the automatic calculation of AHI.

Pediatric sleep apnea: Characterization of apneic events and sleep stages using heart rate variability

Heart rate variability (HRV) is modulated by sleep stages and apneic events. Previous studies in children compared classical HRV parameters during sleep stages between obstructive sleep apnea (OSA) and controls. However, HRV-based characterization incorporating both sleep stages and apneic events has not been conducted. Furthermore, recently proposed novel HRV OSA-specific parameters have not been evaluated. Therefore, the aim of this study was to characterize and compare classic and pediatric OSA-specific HRV parameters while including both sleep stages and apneic events. A total of 1610 electrocardiograms from the Childhood Adenotonsillectomy Trial (CHAT) database were split into 10-min segments to extract HRV parameters. Segments were characterized and grouped by sleep stage (wake, W; non-rapid eye movement, NREMS; and REMS) and presence of apneic events (under 1 apneic event per segment, e/s; 1-5 e/s; 5-10 e/s; and over 10 e/s). NREMS showed significant changes in HRV parameters as apneic event frequency increased, which were less marked in REMS. In both NREMS and REMS, power in BW2, a pediatric OSA-specific frequency domain, allowed for the optimal differentiation among segments. Moreover, in the absence of apneic events, another defined band, BWRes, resulted in best differentiation between sleep stages. The clinical usefulness of segment-based HRV characterization was then confirmed by two ensemble-learning models aimed at estimating apnea-hypopnea index and classifying sleep stages, respectively. We surmise that basal sympathetic activity during REMS may mask apneic events-induced sympathetic excitation, thus highlighting the importance of incorporating sleep stages as well as apneic events when evaluating HRV in pediatric OSA.

Screening the risk of obstructive sleep apnea by utilizing supervised learning techniques based on anthropometric features and snoring events

Objectives

Obstructive sleep apnea (OSA) is typically diagnosed by polysomnography (PSG). However, PSG is time-consuming and has some clinical limitations. This study thus aimed to establish machine learning models to screen for the risk of having moderate-to-severe and severe OSA based on easily acquired features.

Methods

We collected PSG data on 3529 patients from Taiwan and further derived the number of snoring events. Their baseline characteristics and anthropometric measures were obtained, and correlations among the collected variables were investigated. Next, six common supervised machine learning techniques were utilized, including random forest (RF), extreme gradient boosting (XGBoost), k-nearest neighbor (kNN), support vector machine (SVM), logistic regression (LR), and naïve Bayes (NB). First, data were independently separated into a training and validation dataset (80%) and a test dataset (20%). The approach with the highest accuracy in the training and validation phase was employed to classify the test dataset. Next, feature importance was investigated by calculating the Shapley value of every factor, which represented the impact on OSA risk screening.

Results

The RF produced the highest accuracy (of >70%) in the training and validation phase in screening for both OSA severities. Hence, we employed the RF to classify the test dataset, and results showed a 79.32% accuracy for moderate-to-severe OSA and 74.37% accuracy for severe OSA. Snoring events and the visceral fat level were the most and second most essential features of screening for OSA risk.

Conclusions

The established model can be considered for screening for the risk of having moderate-to-severe or severe OSA.

A review of automated sleep disorder detection

Automated sleep disorder detection is challenging because physiological symptoms can vary widely. These variations make it difficult to create effective sleep disorder detection models which support hu-man experts during diagnosis and treatment monitoring. From 2010 to 2021, authors of 95 scientific papers have taken up the challenge of automating sleep disorder detection. This paper provides an expert review of this work. We investigated whether digital technology and Artificial Intelligence (AI) can provide automated diagnosis support for sleep disorders. We followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines during the content discovery phase. We compared the performance of proposed sleep disorder detection methods, involving differ-ent datasets or signals. During the review, we found eight sleep disorders, of which sleep apnea and insomnia were the most studied. These disorders can be diagnosed using several kinds of biomedical signals, such as Electrocardiogram (ECG), Polysomnography (PSG), Electroencephalogram (EEG), Electromyogram (EMG), and snore sound. Subsequently, we established areas of commonality and distinctiveness. Common to all reviewed papers was that AI models were trained and tested with labelled physiological signals. Looking deeper, we discovered that 24 distinct algorithms were used for the detection task. The nature of these algorithms evolved, before 2017 only traditional Machine Learning (ML) was used. From 2018 onward, both ML and Deep Learning (DL) methods were used for sleep disorder detection. The strong emergence of DL algorithms has considerable implications for future detection systems because these algorithms demand significantly more data for training and testing when compared with ML. Based on our review results, we suggest that both type and amount of labelled data is crucial for the design of future sleep disorder detection systems because this will steer the choice of AI algorithm which establishes the desired decision support. As a guiding principle, more labelled data will help to represent the variations in symptoms. DL algorithms can extract information from these larger data quantities more effectively, therefore; we predict that the role of these algorithms will continue to expand.

Environmental Benefits of Sleep Apnoea Detection in the Home Environment

Sleep Apnoea (SA) is a common chronic illness that affects nearly 1 billion people around the world, and the number of patients is rising. SA causes a wide range of psychological and physiological ailments that have detrimental effects on a patient’s wellbeing. The high prevalence and negative health effects make SA a public health problem. Whilst the current gold standard diagnostic procedure, polysomnography (PSG), is reliable, it is resource-expensive and can have a negative impact on sleep quality, as well as the environment. With this study, we focus on the environmental impact that arises from resource utilisation during SA detection, and we propose remote monitoring (RM) as a potential solution that can improve the resource efficiency and reduce travel. By reusing infrastructure technology, such as mobile communication, cloud computing, and artificial intelligence (AI), RM establishes SA detection and diagnosis support services in the home environment. However, there are considerable barriers to a widespread adoption of this technology. To gain a better understanding of the available technology and its associated strength, as well as weaknesses, we reviewed scientific papers that used various strategies for RM-based SA detection. Our review focused on 113 studies that were conducted between 2018 and 2022 and that were listed in Google Scholar. We found that just over 50% of the proposed RM systems incorporated real time signal processing and around 20% of the studies did not report on this important aspect. From an environmental perspective, this is a significant shortcoming, because 30% of the studies were based on measurement devices that must travel whenever the internal buffer is full. The environmental impact of that travel might constitute an additional need for changing from offline to online SA detection in the home environment.

Conventional Machine Learning Methods Applied to the Automatic Diagnosis of Sleep Apnea

The overnight polysomnography shows a range of drawbacks to diagnose obstructive sleep apnea (OSA) that have led to the search for artificial intelligence-based alternatives. Many classic machine learning methods have been already evaluated for this purpose. In this chapter, we show the main approaches found in the scientific literature along with the most used data to develop the models, useful and large easily available databases, and suitable methods to assess performances. In addition, a range of results from selected studies are presented as examples of these methods. Very high diagnostic performances are reported in these results regardless of the approaches taken. This leads us to conclude that conventional machine learning methods are useful techniques to develop new OSA diagnosis simplification proposals and to act as benchmark for other more recent methods such as deep learning.

Oximetry Indices in the Management of Sleep Apnea: From Overnight Minimum Saturation to the Novel Hypoxemia Measures

Obstructive sleep apnea (OSA) is a multidimensional disease often underdiagnosed due to the complexity and unavailability of its standard diagnostic method: the polysomnography. Among the alternative abbreviated tests searching for a compromise between simplicity and accurateness, oximetry is probably the most popular. The blood oxygen saturation (SpO₂) signal is characterized by a near-constant profile in healthy subjects breathing normally, while marked drops (desaturations) are linked to respiratory events. Parameterization of the desaturations has led to a great number of indices of severity assessment commonly used to assist in OSA diagnosis. In this chapter, the main methodologies used to characterize the overnight oximetry profile are reviewed, from visual inspection and simple statistics to complex measures involving signal processing and pattern recognition techniques. We focus on the individual performance of each approach, but also on the complementarity among the great amount of indices existing in the state of the art, looking for the most relevant oximetric feature subset. Finally, a quick overview of SpO₂-based deep learning applications for OSA management is carried out, where the raw oximetry signal is analyzed without previous parameterization. Our research allows us to conclude that all the methodologies (conventional, time, frequency, nonlinear, and hypoxemia-based) demonstrate high ability to provide relevant oximetric indices, but only a reduced set provide non-redundant complementary information leading to a significant performance increase. Finally, although oximetry is a robust tool, greater standardization and prospective validation of the measures derived from complex signal processing techniques are still needed to homogenize interpretation and increase generalizability.