Application of artificial intelligence in the diagnosis of sleep apnea

Abstract

Study Objectives:

Machine learning (ML) models have been employed in the setting of sleep disorders. This review aims to summarize the existing data about the role of ML techniques in the diagnosis, classification, and treatment of sleep-related breathing disorders.

Methods:

A systematic search in Medline, EMBASE, and Cochrane databases through January 2022 was performed.

Results:

Our search strategy revealed 132 studies that were included in the systematic review. Existing data show that ML models have been successfully used for diagnostic purposes. Specifically, ML models showed good performance in diagnosing sleep apnea using easily obtained features from the electrocardiogram, pulse oximetry, and sound signals. Similarly, ML showed good performance for the classification of sleep apnea into obstructive and central categories, as well as predicting apnea severity. Existing data show promising results for the ML-based guided treatment of sleep apnea. Specifically, the prediction of outcomes following surgical treatment and optimization of continuous positive airway pressure therapy can be guided by ML models.

Conclusions:

The adoption and implementation of ML in the field of sleep-related breathing disorders is promising. Advancements in wearable sensor technology and ML models can help clinicians predict, diagnose, and classify sleep apnea more accurately and efficiently.

Citation:

Bazoukis G, Bollepalli SC, Chung CT, et al. Application of artificial intelligence in the diagnosis of sleep apnea. J Clin Sleep Med. 2023;19(7):1337–1363.

INTRODUCTION

Sleep apnea is characterized by repetitive episodes of complete (apnea) or partial (hypopnea) cessation of breathing during sleep. There are two mechanistic categories of sleep apnea - obstructive sleep apnea (OSA), which is characterized by upper airway blockage during sleep with continued respiratory effort, and central sleep apnea, in which both airflow and inspiratory effort are absent or reduced. Sleep apnea is a major health issue associated with different forms of cardiovascular disease including hypertension, coronary artery disease, and stroke.¹ The standard test for the diagnosis of sleep apnea is polysomnography (PSG) during which patients need to sleep overnight at a sleep laboratory. However, more recently home sleep testing is increasingly being used to diagnose OSA. The most commonly used type of home sleep testing device records airflow, respiratory effort, oxygen saturation, and heart rate.

The management of this breathing disorder is related to the mechanism of apnea. The role of machine learning (ML) techniques has been studied in different aspects of medicine, including diagnosis, risk stratification, response to treatment, and personalized management.^2–4 This review aims to present the existing data about the role of ML techniques in the diagnosis, classification, and treatment of sleep apnea.

METHODS

Search strategy

This review aimed to identify the existing studies about the role of ML techniques in sleep apnea. A flowchart of the search strategy is presented in Figure 1. A systematic search in Medline, EMBASE, and Cochrane databases through January 2022 was performed.

Figure 1. Flowchart of the search strategy.

The reference lists of the relevant studies and relevant review studies were manually searched. The following keywords were used in the search strategy: “machine learning,” “artificial intelligence,” and “sleep apnea.”

Inclusion/exclusion criteria

We included studies that provided data about the usage of ML in the management of sleep apnea. Studies that did not provide relevant data, review studies, meta-analyses, and studies written in another language than English were excluded.

Data extraction

The following data were extracted: first author, journal of publication, year of publication, outcomes, machine learning methods, and model input features.

Quality assessment

Assessment for risk of bias and quality were completed using the QUADAS-2 tool.⁵ Two categories, risk of bias and concerns regarding applicability, were assessed in the 3 domains of patient selection, index test, and reference standard, with the former being assessed in addition by the domain flow and timing. For specific assessment of the risk of bias, we have set the following criteria, that is, for each of the 4 domains: (1) when the answer to each question is “yes”, the overall bias risk of the domain is “low”; (2) when the answer to more than one question is “no”, bias risk was definitely identified, and the overall bias risk of the domain is “high”; (3) deemed “unclear” when the data reported is insufficient to make a judgment; (4) when any domain is high risk, the overall bias risk score is “high”; (5) only when the bias risk of 1 domain is unclear, the overall bias risk of the study is “unclear”.

The recommendation of the QUADAS-2 tool was followed, and the clinical applicability of each study was scored by evaluating whether it matched the concerns of our review, and rated as “low”, “high”, or “unclear”. An author (XL) independently performed the data extraction and quality assessment. The final study quality was classified as low risk of bias, high risk of bias, and unclear.

The present study is a systematic review and therefore an ethical approval was not required. A protocol was not prepared for this study and the study was not registered prospectively.

A primer on machine learning

ML is a powerful technology that enables computers to learn directly from the data to draw specific inferences without being explicitly programmed. ML has been extensively used in various domains including health care to analyze large volume of data and make better and more informed decisions. Development of an ML model includes data aggregation and preprocessing, handcrafted feature engineering, model selection and optimization, parameter tuning, model training, model validation, and model deployment (Figure 2). It is an iterative process where various ML models are fine-tuned using training data and evaluated using validation data to arrive at an optimal model. The performance of an optimal model is evaluated using the unseen test data. Various ML algorithms have been developed for diagnosing sleep apnea. Below, we provide a brief introduction to some of the ML algorithms we encounter in this manuscript.

Figure 2. A machine learning pipeline.

An iterative process to obtain an optimal model using training and validation data is displayed; finally, the optimal model is evaluated on unseen test data.

A linear regression model obtains a linear mapping from input variables to a continuous output variable.⁶ A logistic regression model is used to predict a binary output variable by mapping the input variable to the probability of the output variable being 1 of 2 classes.⁷ Decision trees are used for both regression and classification tasks by building a tree-like structure with each node comparing a feature to a threshold value and each leaf representing the predicted outcome.⁸ Threshold values at each node are optimized to obtain the desired outcome. Output from multiple decision trees are considered to create random forest models.^9,10 K-nearest neighbors (kNN) is an instance-based learning algorithm; an instance’s classification or regression value is determined by the majority class or average value, respectively, of the k-nearest instances.¹¹ Support vector machine (SVM) algorithm finds a linear boundary (hyperplane) that separates the features into different classes with maximum margin.¹² A nonlinear hyperplane can be obtained by using a kernel-based SVM that transform the data into a higher-dimensional space, where a linear hyperplane can classify the data.^13,14 Kernel SVMs are effective in classifying high-dimensional and nonlinear data.

Neural networks are computational models inspired by the way biological neural networks in the human brain process information.¹⁵ They are made up of layers of interconnected nodes, called artificial neurons, where each layer applies a set of weights and biases followed by nonlinear activation. Such parameters are tuned to learn the complex, nonlinear relationships between inputs and outputs. Based on the connectivity across different layers, there are various types of neural networks. Multilayer perceptron is feedforward neural network consisting of an input layer fully connected to one or more hidden layers, and an output layer.¹⁶ A convolutional neural network (CNN) consists of 1 or more convolutional layers that apply a convolution operation on the data.¹⁷ Such convolution operations automatically extract the features by capturing the inherent spatial and temporal dependencies within the data. Recurrent neural networks have a feedback loop within the network to maintain an internal state that can be updated over time, which allows them to process sequential data.^18,19 A bidirectional recurrent neural networks processes the sequential input data from both directions, using two separate hidden layers.²⁰ Long short-term memory network and gated recurrent unit are recurrent neural networks that use specific memory cells and gates to control the flow of information within the network to capture long-term dependencies.^21,22

RESULTS

The search strategy revealed 498 studies: 304 studies were excluded at the title/abstract level and 62 were excluded at the full-text level. Finally, 132 studies were included in the systematic review. Of them, 81 studies reported data about the identification/diagnosis of sleep apnea, 19 studies about the prediction of sleep apnea, 17 studies about the classification of sleep apnea, and 15 studies provided data for different outcomes not classified elsewhere (Figure 1 and Table 1).

Table 1.

Main characteristics and provided outcomes of the included studies.

Study	Year	Outcome / Definition of OSA in “diagnosis of sleep apnea” studies	Features / sleep time / Desaturation cut-off	ML approach used	Comparison between ML vs conventional tools	AUC	Specificity	Sensitivity	Accuracy	PPV	NPV
Diagnosis of Sleep Apnea
Alvarez-Estevez et al46	2016	Screening of SAHS patients AHI ≥ 5%	Heart timing VS heart rate and heart period Sleep time: 392 minutes	Naive Bayes classifier		0.88	1.00	0.75	0.83	—	—
Sharma and Sharma47	2016	Diagnosis of sleep apnea AHI ≥ 5%	Hermite decomposition of the QRS complex (single-lead ECG)	KNN, MLPNN, SVM, and LS-SVM	LDA	—	—	—	1.00	—	—
					SVM, LS-SVM classifier	—	—	—	0.971	—	—
Viswabhargav et al24	2019	Diagnosis of sleep apnea AHI ≥ 5%	Sparse residual entropy using ECG-derived respiration and heart rate signals	fuzzy K-means clustering and SVM	Fourier transform-based CP features (Threshold)	—	0.68	0.81	—	—	—
					Radial basis function kernel-based SVM classifier	—	0.781	0.78	0.781	—	—
Mashrur et al36	2021	Diagnosis of OSA AHI ≥ 5%	Single-lead ECG	SCNN	SCNN	—	0.945	0.943	0.944	—	—
Babaeizadeh et al26	2010	Diagnosis and quantification of sleep apnea (OSA or mixed) AHI 5%	ECG data Low-frequency fluctuations of instantaneous HR	Quadratic classifier		—	1.00	0.95	0.97	—	—
Marcos et al76	2008	Diagnosis of OSA AHI ≥ 10%	Nonlinear features from nocturnal oxygen saturation	MLP neural networks		—	0.794	0.898	0.855	—	—
Del Campo et al53	2006	Diagnosis of OSA AHI ≥ 10%	Arterial oxygen saturation (Desaturation cut-off 2%, 3% and 4%) Frequency of 0.2 Hz	Approximate entropy analysis	Total time spent with SaO2 below 90% (SaO2 > 1%)	0.774	0.755	0.73	—	0.78	0.702
					Approximate entropy analysis	0.921	0.829	0.833	—	0.883	0.829
Aydoğan et al90	2016	Diagnosis of OSA AHI ≥ 5%	Respiratory signals	Morphological filter via ANN	Visual scoring-based algorithm	—	—	—	0.883	—	—
					Morphological filter via ANN	—	—	—	0.873	—	—
Weinreich et al164	2008	Diagnosis of OSA and Cheyne-Stokes respiration AHI ≥ 10%	Flow-related spectral entropy	ANN	n/a	0.988	0.946	0.913	0.915	—	—
Ma et al78	2020	Diagnosis of sleep apnea	Blood oxygen saturation	SVM	AdaBoost	—	0.872	0.727	0.836	—	—
					SVM	—	0.941	0.876	0.902	—	—
Wang et al25	2019	Diagnosis of sleep apnea AHI ≥ 5%	RR intervals amplitudes of a single ECG	LeNet-5, MLP, KNN, LR, SVM	LeNet-5t	—	0.917	1.00	0.971	—	—
Liu et al51	2008	Diagnosis of sleep apnea	Pupil size, electroencephalogram	ANN		—	—	—	0.91	—	—
Al-Abed et al35	2006	Diagnosis of OSA	HRV	Fuzzy classification		—	0.685	0.832	0.759	—	—
Aksahin et al165	2012	Diagnosis and classification of sleep apnea	Electroencephalogram	Feedforward Neural Network, Radial Basis Function NN and Distributed Time Delay Neural Network		—	—	—	—	—	—
Alshaer et al79	2019	Identification of sleep apnea AHI 5%, 10%, and 15% cut-off were used	Snoring frequency	Advanced signal processing algorithms		—	0.31	0.91	0.975	0.63	0.62
Alvarez et al54	2020	Diagnosis of OSA AHI 5%, 15%, and 30% cut-off were used	Oximetry analysis, airflow recordings	Regression SVM		0.98	0.98	0.936	0.958	0.922	0.842
Hajipour et al138	2020	Diagnosis of OSA AHI ≥ 5%	Daytime tracheal breathing sounds	RF	LR	—	0.758	0.822	0.793	—	—
					RF	—	0.795	0.842	0.821	—	—
Acharya et al48	2011	Diagnosis of sleep apnea/hypopnea AHI ≥ 5%	ECG signals	ANN		—	1.00	0.95	0.9	—	—
Zhou et al49	2015	Automatic detection of sleep apnea AHI ≥ 5%	Electroencephalogram signals	SVM		—	0.986	0.932	0.951	—	—
Behar et al60	2019	Idenification of OSA AHI ≥ 5%	All oxygen saturation and the demographic features available from the STOP-BANG questionnaire (desaturation cut-off: 3%)	Random subsampling validation	Logistic regression model (OxyDOSA)	0.94	0.85	0.87	0.86	0.82	0.9
					STOP-BANG questionnaire	0.77	0.81	0.61	0.72	0.7	0.73
Bernardini et al75	2021	Identification of OSA in stroke patients AHI ≥ 5%	ECG and SpO2	CNN + LSTM		0.825	0.843	0.672	0.815	—	—
Bertoni et al101	2020	Identification of obstructive sleep-disordered breathing in children postoperatively AHI cut offs 2% and 10%	Clinical parameters and nocturnal actigraphy and oxygen desaturation index	Recursive partitioning, conditional inference, RF, bootstrap aggregation, SVM, logistic regression	Recursive partitioning	—	0.88	0.90	0.89	0.93	0.82
					Actigraphy	—	0.46	0.88	0.84	—	—
Bricout et al70	2021	Sleep-apnea screening AHI > 15	CONF polysomnography CONF accelerometry derived respiratory index	Boosted tree	ADR	—	1.00	0.8	0.89	—	—
					Polysomnography	—	1.00	1.00	1.00	—	—
Calderon et al95	2020	Identification of children at risk of OSA AHI ≥ 5%	Oxygen saturation data	SVM, LR, AdaBoost	LR	0.9	0.96	0.62	0.79	0.94	—
					Multivariate adaptive regression splines	—	0.54	0.83	—	—	—
Chang et al139	2020	Screening of OSAHS Nasal flow cessation > 10 seconds plus paradoxical respiratory and abdominal efforts	LSTM-NN Sleep time: 307.9 minute	SpO2, ECG, and thoracic triaxial accelerometers signals	LSTM-RNN	—	—	0.81	0.893	—	—
Chang et al27	2020	Per minute apnea detection AHI ≥ 5%	Deep CNN	Single-lead ECG signal	Auto-encoder + Decision Fusion	0.87	0.821	0.889	0.847	—	—
					Deep CNN	0.94	0.92	0.811	0.879	—	—
De Silva et al80	2011	Diagnosis of OSA AHI cut offs 5%, 15% and 30%	Snore signals	NN		—	0.89	0.91	—	—	—
ElMoaqet et al72	2020	Diagnosis of sleep apnea	Single channel respiratory signals	RNN		0.924	0.837	0.903	0.85	0.588	0.971
Erdenebayar et al28	2019	Diagnosis of sleep apnea events AHI ≥ 5%	ECG signal	Deep neural network, 1-dimensional CNN, 2-dimensional CNN, RNN, LSTM, and gated recurrent unit	GRU	—	0.99	0.99	0.99	—	—
Ganglberger et al71	2021	Diagnosis of sleep apnea AHI ≥ 5%	Respiratory and SpO2 signals	RF		0.94	—	0.58	0.94	—	—
Gao et al91	2019	Diagnosis of OSA	HRV, Respiratory signals	Decision Tree classification and model fusion	Decision tree classification and model fusion	—	0.74	0.74	0.75	—	—
					ECG quality exclusion criteria	—	0.857	0.625	0.733	—	—
Barroso-Garcia et al96	2020	Diagnosis of pediatric sleep apnea AHI ≥ 1%	Airflow recordings- recurrence plots features	MLP-NN with Bayesian approach		—	0.943	0.788	0.91	0.788	0.943
Xie et al64	2021	Screening and diagnosing OSA	Audio recordings	CNN, RNN	CNN	—	0.977	0.922	0.953	—	—
Xu et al97	2019	Diagnosis of OSA in symptomatic snoring children AHI ≥ 1%	Signal from oximetry measurements Sleep time: 474.1 minutes	MLP		—	0.927	0.735	0.882	—	—
Hang et al55	2015	Diagnosis of moderate to severe OSA AHI ≥ 5%	Oximetry measurements Apnea definition: minimum of 10 seconds of airflow cessation Hypopnea definition: 30% reduction in airflow preceding a period of normal breathing for a minimum of 10 seconds and oxyhemoglobin desaturation (decrease in SpO2 ≧ 4%) Frequency of 1 Hz	SVM		0.924	0.931	0.899	0.906	—	—
Andres-Blanco et al56	2017	Diagnosis of OSA AHI ≥ 5%	SpO2 signals (desaturation cut-off 3% and 4%)	MLP		—	0.586	0.975	0.873	0.868	0.895
Pradhapan et al73	2013	Identification of apnea	PPG signals–frequency spectrum analysis	SVM		—	0.989	0.950	0.968	—	—
Hwang et al65	2015	Snoring detection in OSA patients	Snoring signals Sleep time: 8.1 hours	SVM	SVM	—	—	0.946	—	0.975	—
					Manual annotations of microphone sensor		—	0.822	—	—	—
Iwasaki et al29	2021	Diagnosis of moderate to severe sleep apnea AHI ≥ 15%	ECG–RR intervals	RNN	RNN	—	1.00	1.00	—	—	—
					Healthdyne 202-11 Oximeter	—	0.8	0.97	—	—	—
De Groote et al98	2002	Detection of obstructive apnea events in infants	Thoracoabdominal signals	ANN		—	0.062	0.75	—	—	—
Jiang et al67	2021	Identification of OSA patients AHI ≥ 5%	Snoring sounds	LR, SVM, Gaussian Bayesian, KNN, ANN	LR	1.0	1.0	1.0	1.0	—	—
Erdenebayar et al81	2017	Detection of OSA events AHI ≥ 5%	Pulse rate variability data Sleep time: 5.83 hours	SVM		—	0.961	0.885	0.956	—	—
Drzazga et al94	2021	Detection of apnea/hypopnea episodes AHI ≥ 5%, 15% or 30% cut-offs	Oronasal airflow, thoracic and abdominal respiratory effort signals	LSTM		—	—	—	0.820	—	—
Shen et al68	2020	Detection of OSAHS AHI ≥ 5%	Snoring signals	CNN, LSTM	MFCC + LSTM	—	0.91	0.84	0.87	—	—
Tiron et al66	2020	Screening for OSA AHI ≥ 5%	Sound signals	NN	“Firefly”	0.92	0.8	0.883	0.842	0.815	0.873
					“ResApp”	0.91	0.83	0.86	—	—	—
Koley et al61	2014	Diagnosis of apnea/hypopnea events AHI 5 or 15%	SpO2 signals (desaturation cut-off: 2%, 3% and 4%)	SVM		—	0.848	0.903	0.967	—	—
Leino et al59	2021	Screening of sleep apnea in patients with Acute cerebrovascular events AHI ≥ 5%	Nocturnal SpO2 (desaturation cut-off: 4%)	CNN		—	0.786	0.918	0.779	—	—
Karamanli et al87	2016	Diagnosis of OSA AHI ≥ 10%	Sex, age, BMI, and snoring status	ANN, MLP	MLP	—	—	—	0.866	—	—
Li et al30	2020	Diagnosis of sleep apnea AHI ≥ 5%	Multiple biosignals	SVM with Linear kernel		0.952	0.962	0.943	0.952	—	—
Li et al44	2021	Diagnosis of OSA AHI ≥ 5%	ECG, pulse oxygen saturation, and BMI	Multilayer feedforward NN		0.97	0.939	0.986	0.978	—	—
Mosquera-Lopez et al82	2019	Diagnosis and classification of OSA AHI ≥ 5%	Pressure sensors in bed at home	Stage-one DT Stage-two linear regression model		—	0.765	0.889	0.743	—	—
Luo et al63	2020	Diagnosis of OSA AHI ≥ 15% or 30%	Sleep sound signals	CNN		0.81	—	—	0.8163	—	—
Lweesy et al39	2011	Diagnosis of OSA	ECG signals P-wave duration, the P-wave dispersion, and the time interval from the peak of the P-wave to the R-wave	ANN		—	—	—	0.923	—	—
Li et al52	2018	Sleep-disordered breathing screening tool AHI ≥ 5%, 10%, 15%, 20%, 25% or 30% cut-off	Nocturnal pulse oximetry with demographic, anatomic, and clinical data (desaturation cut-off: 4%)	ANN		0.954	0.716	0.983	—	0.286	0.997
Marcos et al58	2010	Diagnosis of OSA AHI ≥ 15%	SpO2 signals (Desaturation cut-off: 2%, 3% and 4%)	Maximum likelihood and Bayesian MLP networks		0.9	0.824	0.878	0.856	—	—
Martinot et al99	2021	Diagnosis of pediatric sleep apnea AHI ≥ 1%	Mandibular movements Sleep time: 7.6 hours	LR		0.98	0.88	1.00	0.94	0.89	1.00
Morillo et al57	2013	Diagnosis of SAHS AHI ≥ 10%	Overnight SpO2 Apnea definition: Reduction in airflow ≥ 90% which was lasting ≥ 10 seconds. Hypopnea definition: Reduction in airflow ≥ 50% accompanied by a decrease of more than 4% in the saturation of hemoglobin	Probabilistic NN	PNN	0.961	0.959	0.924	0.939	—	—
					Univariate approach (RDI > 15 Threshold)	—	0.75	0.9	—	—	—
Mukherjee et al38	2021	Diagnosis of sleep apnea AHI ≥ 5%	ECG signals	CNN, CNN-LSTM	MLP	—	0.883	—	0.856	—	—
Nakano et al69	2019	Diagnosis of sleep apnea AHI ≥ 5%	Tracheal sound analysis	DNN		—	0.72	0.92	0.88	0.93	—
Al-Angari et al83	2012	Recognition of OSA AHI ≥ 5%	Respiratory features, oxygen saturation, heart rate variability	SVM	Polynomial kernel	—	0.98	0.918	0.95	—	—
Niroshana et al42	2021	Diagnosis of OSA AHI ≥ 5%	Single-lead ECG data	CNN	CNN	—	0.926	0.923	0.924	—	—
					Cardio-pulmonary signal/fast and adaptive bivariate EMD coupled with cross time-frequency analysis	—	0.787	0.823	—	—	—
Nakayama et al34	2019	Apnea/normal respiration discrimination AHI ≥ 15%	HRV	RF		0.84	0.92	0.76	—	—	—
Ostadieh et al43	2020	Diagnosis of OSA	ECG signals	Hybrid “K‐Means, Recursive Least‐Squares” Learning for the Radial Basis Function Network	Hybrid “k-means, RLS” RBF	—	0.96	0.964	0.956	—	—
					Statistical analysis	—	0.81	1.00	0.93	—	—
Papini et al93	2020	Estimation of AHI AHI ≥ 5%, 15% or 30% cut offs	Wrist‐worn reflective photoplethysmography Sleep time: 411 minutes	Deep learning		0.8	0.98	0.41	—	0.8	—
Xie et al62	2012	Diagnosis of sleep apnea	ECG and SpO2 signals (desaturation cut-off: 4%)	AdaBoost with Decision Stump, Bagging with REPTree, kNN, Decision Table, MLP, SVM, FT trees, C4.5 tree	Bagging. REPTree	—	0.846	0.791	0.833	—	—
Ryu et al84	2021	Diagnosis of OSA AHI ≥ 5%	Aerodynamic and biometric features	3D UNet deep-learning model, SVM	SVM	—	0.862	0.893	0.815	—	—
					Berlin questionnaire STOP STOP-Bang questionnaire	—	0.571	0.976	—	—	—
Song et al40	2016	Diagnosis of OSA AHI ≥ 5%	ECG signals	Discriminative hidden Markov model	HMM + SVM	1.00	1.00	0.958	0.971	—	—
					Visual inspection	—	1.00	1.00	1.00	—	—
Khandoker et al41	2009	Automated recognition of OSA AHI ≥ 10%	ECG signals	SVM		—	—	—	0.929	—	—
Moret-Bonillo et al166	2014	Diagnosis of SAHS	Respiratory signals, SpO2	Fuzzy logic		0.864	0.916	0.813	—	—	—
Tripathy et al74	2020	Automated sleep apnea detection	ECG signals	SVM, RF	SVM	—	0.787	0.823	0.783	—	—
Tsuiki et al92	2021	Diagnosis of OSA AHI ≥ 5%	2D lateral cephalometric radiographs	Deep CNN		0.92	0.77	0.9	—	0.87	0.82
Tuncer et al86	2019	Diagnosis of OSA AHI ≥ 5%	Pulse transition Ttime	CNN, kNN, SVM, Alex-Net, VGG-16	SVM	—	0.98	—	0.9278	—	—
Urtnasan et al45	2018	Diagnosis of sleep apnea events AHI ≥ 5%	Single-lead ECG signals Sleep time: 5,72 hours	CNN	CNN	—	0.96	0.96	0.96	—	—
					Kernel density classifier	—	0.802	0.832	0.821	—	—
Varon et al31	2015	Diagnosis of sleep apnea AHI ≥ 10%	Single-lead ECG	least-squares support vector machines, SVM, LDA	LV-SVM RBF	0.9	0.846	0.788	0.840	—	—
Vaquerizo-Villar et al100	2018	Screening of pediatric sleep AHI AHI ≥ 5%	Bispectrum-based oximetry Sleep time: 7,8 hours	MLP-NN		—	0.945	0.6	0.853	0.8	0.867
Vimala et al50	2019	Diagnosis of sleep apnea	Electroencepholgraphy signals	SVM, kNN, ANN	SVM	—	0.98	1.00	0.99	—	—
Wang et al32	2019	Diagnosis of apnea	ECG signals (RR intervals)	Deep residual network, CNN	Residual network	—	0.95	0.93	0.944	—	—
					Frequency network	—	0.905	0.883	0.901	—	—
Wang et al33	2019	Diagnosis of sleep apnea AHI ≥ 5%	Single-lead ECG signal	Time window-MLP, MLP, SVM, LDA, LR	TW-MLP	—	0.887	0.851	0.873	—	—
Yue et al88	2021	Diagnosis and classification of sleep apnea AHI ≥ 5%	Nasal airflow signals OSA event: ≥ 90% drop in respiratory flow for at least 10 seconds, associated with breathing effort. Hypopnea event: ≥ 30% drop in respiratory flow for at least 10 seconds, associated with ≥ 3% oxygen desaturation or arousal.	Residual network		—	0.905	0.908	0.912	—	—
Zhang et al37	2021	Diagnosis of OSA	Single-channel ECG	Deep CNN-LSTM		—	0.962	0.961	0.961	0.976	0.938
Zhang et al89	2021	Screening of moderate to severe OSA AHI ≥ 15%	Faciocervical and anthropometric measurements	SVM	Sex-Age- BMI-maximum interincisal distance-ratio of height to thyrosternum distance-neck circumference-waist circumference (SABIHC2) model	0.832	0.749	0.916	—	—	—
					STOP-BANG questionnaire	0.631	0.772	0.487	—	—	—
Prediction of Sleep Apnea
Taghizadegan et al102	2021	Prediction of OSA	Polysomnography signals (electroencephalogram and ECG)	RPCNNs		0.911	0.915	0.916	0.917	—	—
Juang et al77	2021	Prediction of OSAHS	Waist circumference, mean BP at the end of polysomnography and the difference in systolic BP between the end and start of polysomnography Sleep time: 331.5 minutes	Explainable fuzzy NN, a back propagation NN and a stepwise regression model		—	0.876	0.573	0.8	—	—
Caffo et al103	2010	Prediction of mild obstructive sleep disordered breathing	Neck circumference, BMI, age, snoring frequency, waist circumference, and snoring loudness	Boosting, RF, NN, LR	Boosting	0.747	—	—	—	—	—
Ferre et al117	2019	Prediction of sleep-related breathing disorders in patients with Chiari malformation type 1	Clinical and neuroradiological parameters	MLR and URP-CTREE	MLR	0.74	0.57	0.92	0.8	0.81	0.77
Huang et al114	2020	Prediction of OSA	Symptoms suggestive of OSA, habitual sleep pattern, comorbidity, AHI, hypnoic, alcohol consumption, smoking status, waist, neck circumference, BMI, sex, age	SVM, LR	SVM	0.78	0.703	0.703	0.703	A0.619	0.779
					NoSAS Score	0.68	0.706	0.649	0.683	0.602	0.746
Keshavarz et al104	2020	Predicition of OSA	Severe snore, nocturia, awakening due to the sound of snoring, witnessed snore, witnessed apnea, back pain, restless sleep, and BMI	SVM, Naïve Bayes, LR, NN, KNN, RF	Naïve Bayes	0.768	0.595	0.790	0.723	—	—
Mencar et al115	2020	Prediction of OSA severity	Demographic characteristics, spirometry values, gas exchange (PaO2, PaCO2) and symptoms	SVM, RF	SVM	0.65	—	0.447	0.447	—	—
Waxman et al118	2010	Prediction of apnea and hypopnea	Electroencephalogram, HRV, nasal pressure, oronasal temperature, submental EMG, and electrooculography	Large Memory Storage and Retrieval		—	0.985	0.886	—	0.858	0.883
Schwartz et al105	2020	Prediction of OSA	Deep sleep questionnaire, sleep parameters	ElasticNet algorithm		0.85	0.665	0.834	0.732	—	—
Kim et al106	2021	Prediction of OSA	Hypertension, waist circumference, length between the subnasale and stomion (subnasale to stomion), snoring from the BQ, loudness of snoring from the BQ, frequency of falling asleep (falling asleep from the BQ), and the FSS total score	LR, SVM, RF, XGBoost	SVM	0.87	0.870	0.803	0.833	0.8852	0.6957
Kim et al167	2019	Prediction of OSA	Respiratory sounds Sleep time: 369.7 minutes	LR		0.91	0.941	0.6	0.853	—	—
Kim et al116	2020	Prediction of AHI	Breathing sounds Sleep time: 369.7 minutes	Gaussian process, SVM, RF, simple linear regression	RF	—	—	—	0.833	—	—
Kirby et al107	1999	Prediction of OSA	Clinical variables	Generalized regression NN		0.94	0.8	0.989	0.913	0.881	0.98
Mihaicuta et al108	2017	OSAS phenotyping and severity prediction	Demographic and clinical characteristics	Classification tree, computational framework		—	0.4189	0.8025	0.693	—	—
Liu et al109	2017	Prediction of the severity of OSA	Anthropometric features	SVM		0.81	0.817	0.664	0.735	—	—
Skotko et al110	2017	Prediction of OSA in patients with Down syndrome	Survey questions, medication history, anthropometric measurements, vital signs, patient's age, and physical examination findings	Logistic learning machine		—	—	—	—	0.25	0.9
El-Solh et al111	1999	Prediction of AHI	Anthropomorphic measurements and clinical information	ANN, multiple linear regression, regression tree	NN	—	0.734	0.955	—	0.833	0.921
Teferra et al113	2014	Prediction of OSA	Demographics, clinical history, questionnaire	ANN	OSUNet	—	0.7	0.63	—	0.67	0.66
					STOP-BANG	—	0.12	0.97	—	0.51	0.82
Ustun et al112	2016	Prediction for sleep apnea	Medical history features	Supersparse Linear Integer Model	SLIM	—	0.77	0.642	—	—	—
					STOP-BANG	—	0.564	0.836	—	—	—
Classification of Sleep Apnea
Fontenla-Romero et al119	2005	Classification of sleep apnea (obstructive, central and mixed)	Thoracic effort signals	NN		—	—	—	0.838	—	—
Baty et al122	2020	Classification of sleep apnea severity	ECG data - HRV	SVM		—	0.74	0.7	0.72	—	—
Urtnasan et al125	2018	Classification of obstructive sleep apnea/hypopnea	Single-lead ECG recordings	CNN		—	0.87	0.87	0.908	—	—
Behar J et al126	2015	Classification of OSA (moderate-severe OSA or healthy)	Audio, actigraphy, photoplethysmography, and demographics	SVM	Audio + Oxygen desaturation index + Actigraphy SVM	0.942	1.00	0.818	0.884	—	—
Rachim et al127	2014	Sleep apnea classification (normal-apnea subjects)	ECG signals	SVM		—	0.952	0.927	0.943	—	—
Park et al128	2015	Classification of apnea-hypopnea events	PPG and SpO2 signals Sleep time: 5.8 hours	SVM	SVM	—	—	0.924	—	0.928	—
					DAP detector	—	—	0.76	—	0.73	—
Wang et al123	2016	OSA severity	Anthropometric and questionnaire data	Fuzzy classification system, FDT, SMOTE, LR, ANN, SVM	SMOTE + FDT	—	—	—	0.8182	—	—
Ding et al129	2021	Severity evaluation	Speech features	SVM		—	0.803	0.773	0.788	—	—
Huttunen et al130	2021	Classification of OSA severity	PPG signals Sleep time: 309 minutes	CNN		—	—	—	0.833	—	—
Jarchi et al131	2020	Classification into normal, OSA, restless leg syndrome (RLS), OSA and RLS	Electromyography and ECG signals	Multimodal deep learning NN		—	—	—	0.72	—	—
Urtnasan et al124	2020	Identification of sleep apnea severity	Single-lead ECG Sleep time: 5,7 hours	CNN		—	—	—	0.99	—	—
Kang et al132	2020	Classification of sleep apnea/hypopnea events	PPG signals	LSTM		—	—	0.86	—	0.942	—
Kim et al133	2018	Classification based on AHI	Breathing sounds	Simple logistics, SVM, and deep neural networks		—	—	—	0.925	—	—
Tagluk et al120	2010	Classification of sleep apnea	Abdominal effort signals	ANN		—	—	—	0.856	—	—
Nikkonen et al134	2019	Classification of OSA severity	SpO2 signals	ANN	ANN (AHI)	—	—	—	90.9	—	—
Nikkonen et al135	2021	Classification of OSA severity	Peripheral blood oxygen saturation, thermistor-airflow, nasal pressure -airflow, and thorax respiratory effort	Long-short term memory NN		—	0.978	0.818	0.87	—	—
McClure et al136	2020	Classification and detection of breathing patterns	Time series of accelerometer and gyroscope readings	CNN		0.94	—	—	0.86	—	—
Different Outcomes
Khandoker et al137	2009	Scoring of OSA/hypopnea events	ECG signals	SVM + Linear kernel		0.97	0.944	1.00	0.988	—	—
El Solh et al141	2007	Assessment of the optimal CPAP pressure	Age, sex, BMI, neck circumference, and baseline AHI Sleep time: 361 minutes	ANN	Regression equation	—	—	—	0.4	—	—
Bozkurt et al85	2019	Respiratory scoring	PPG signal and HRV	kNN, multilayer ANN, probabilistic ANN, SVM, ensemble classifier	Ensemble classifier	—	0.96	0.93	0.95	—	—
El Solh Ali et al142	2009	CPAP titration	Age, sex, BMI, neck circumference, and AHI	ANN		—	—	—	—	—	—
Elwali et al23	2021	Prediction of PSG parameters Patients referred to PSG study	Breathing sounds and anthropometric features	RF		—	—	—	0.888	—	—
Remmers et al144	2017	Identification of OSA patients who will respond to oral appliance therapy	Data from feedback-controlled mandibular positioner	RF		—	0.93	0.85	—	0.97	0.72
Kim et al145	2021	Prediction of therapeutic outcome of sleep surgery in OSA	Demographic, anatomical parameters, preoperative PSG	LR, RF, SVM, grading boosting machine	Gradient boosting	0.727	0.708	0.708	0.708	0.708	0.708
					Physician’s prediction	—	0.795	0.238	0.522	0.528	0.520
Maranate et al149	2015	Prioritization of clinical risk factors	Clinical risk factors	Fuzzy Analytic Hierarchy Process		—	0.918	0.923	—	—	—
Prasad et al150	2020	Prediction of nocturnal BP in OSA	Clinical PSG signals Sleep time: 333.8 minutes	Deep NN		—	—	—	—	—	—
Rafael-Palou et al143	2018	Assessment of compliance with CPAP therapy	Clinical history, symptoms, comorbidities, therapies, sleep testing	RF, LR		—	—	—	—	—	—
Sebastian et al147	2020	Prediction of predominant site of upper airway collapse	Mel frequency cepstral coefficients	LDA		—	—	—	0.72	—	—
Sebastian et al148	2021	Prediction of predominant site of upper airway collapse	Snore signals	LDA, cluster analysis	LDA	—	0.8	0.83	0.81	—	—
Stretch et al151	2019	Identification of patients with nondiagnostic home sleep apnea tests	Patient responses, measurement, comorbidities, demographics, measurements	RF, ANN, LASSO regression, SVM, kNN, gradient boosted decision tree, ridge regression	RF	—	0.93	0.51	—	0.76	0.81
Waxman et al152	2015	Prediction of disordered breathing events in OSA	HRV, electroencephalogram, ECG, oronasal temperature, nasal pressure, submental electromyography	Large Memory Storage and Retrieval artificial neural networks		—	0.77	0.81	—	—	—
Yang et al146	2021	Prediction of outcomes in palatal surgery for OSA	Septoplasty or functional endoscopic sinus surgery, palate surgery: uvulopalatal flap or palatal muscle resection, demography, polysomnography, Friedman stage, drug-induced sleep endoscopy	LR, tree-based models, bagging, RF, SVM, NN	Lasso	0.832	—	—	—	—	—

ADR = accelerometry-derived respiratory index, AHI = apnea-hypopnea index, ANN = artificial neural network, AUC = area under the curve, BMI = body mass index, BP = blood pressure, BQ = Berlin questionnaire, CNN = convolutional neural network, CONF = configuration, CP = cardiopulmonary, DT = decision tree, ECG = electrocardiogram, EMG = electromyogram, EMD = empirical mode decomposition, FDT = fuzzy decision tree, HRV = heart rate variability, KNN = k-nearest neighbor, kNN = k-nearest neighbor, LASSO = Least Absolute Shrinkage and Selection Operator, LDA = Linear discriminant analysis, LS-SVM = least-square support vector machine, LSTM = long short-term memory, ML = machine learning, MLPNN = multilayer perceptron neural network, MLR = multiple logistic regression, NN = neural network, NPV = negative predictive value, OSA = obstructive sleep apnea, OSAHS = obstructive sleep apnea hypopnea syndrome, PNN = perceptron neural network, PPV = positive predictive value, RDI = Respiratory disturbance index, RF = random forest, RNN = recurrent neural network, RPCNN = recurrence plot convolutional neural networks, SCNN = scalogram-based convolutional neural network, SMOTE = synthetic minority over-sampling technique, SVM = support vector machine, TW-MLP = Multilayer perceptron, URP-CTREE = the unbiased recursive partitioning technique conditional inference tree, VS = versus.

Diagnosis and screening of sleep apnea

Electrocardiogram-based diagnosis

An important role of ML models in sleep apnea is to aid the screening of high-risk populations and facilitate sleep apnea diagnosis. For that matter, ML algorithms have been designed and tested to predict PSG parameters. Specifically, in this setting, a 2-class random forest (RF) classifier resulted in a classification accuracy of up to 88.8% in predicting key PSG parameters using anthropometric features and breathing sounds.²³ In another study, the authors proposed the novel sparse residual entropy features for automated sleep apnea detection using electrocardiogram (ECG)-derived respiration and heart rate signals.²⁴ The experimental results demonstrated that the proposed features with radial basis function kernel-based support vector machine (SVM) classifier yielded higher performance with an accuracy of 78.07% with Fourier dictionary and 10-fold cross-validation, compared to other dictionaries (Cosine dictionary, Eigen dictionary, Wavelet dictionary, Normal-specific learned dictionary, and Apnea-specific learned dictionary).²⁴

A convolutional neural network (CNN) has also shown good performance in diagnosing sleep apnea using ECG recordings.²⁵ Babaeizadeh et al²⁶ proposed a sleep apnea detection algorithm based on analysis of a single ECG lead, which showed that accurate apnea detection and quantification could be achieved. A 1-dimensional deep CNN model using single-lead ECG demonstrated 97.1% accuracy, 100% specificity, and 95.7% sensitivity for per-recording classification, outperforming several feature-engineering-based and feature-learning-based approaches.²⁷ Similarly, Erdenebayar et al²⁸ showed that deep learning approaches such as 1-dimensional CNN and gated recurrent unit could aid the automatic detection of sleep apnea events using single ECG signals. Application of a long short-term memory model on RR interval resulted in sensitivity and specificity of 100% in apnea detection.²⁹ Electrocardiography, oxygen saturation (SpO₂), airflow, abdominal, and thoracic signals used as inputs in an SVM linear kernel, outperformed other classifiers for the detection of sleep apnea.³⁰

An automated sleep apnea detection approach that used a single-lead ECG data, achieved accuracies of ∼85% on a minute-by-minute basis for 2 independent datasets including both hypopneas and apneas. In addition, discrimination between apnea and normal recordings was achieved with 100% accuracy.³¹

A CNN using ECG signals showed accuracy, sensitivity, and specificity for detecting sleep apnea of 94.4%, 93%, and 94.9%, respectively.³² Wang et al³³ developed a time window artificial neural network (ANN) using a single-lead ECG signal for detecting sleep apnea. The proposed model achieved an accuracy of 87.3%, outperforming other traditional ML models such as logistic regression (LR), linear discriminant analysis, SVM, and multilayer perceptron. An RF was used for apnea detection using HRV data, and this model showed a sensitivity and specificity of 76% and 92%.³⁴

In the setting of OSA diagnosis, fuzzy classification of time-frequency heart rate variability (HRV) plots has been shown to detect OSA³⁵ with an accuracy of 75,88%. Mashrur et al³⁶ proposed a novel scalogram-based CNN to detect OSA using single-lead ECG signals; this model showed an accuracy of 94.30% per segment and 100% per-recording classification, outperforming the existing OSA detection approaches using ECG signals.³⁶ Application of CNN on single-channel ECG by portable OSA monitor devices showed a sensitivity, specificity, and accuracy of 96.1%, 96.2%, and 96.1%, respectively.³⁷ ECG data has been shown to accurately to detect OSA using the multilayer perceptron-based ensemble approach with an accuracy of 85.58%.³⁸ ECG and especially P-wave features were analyzed with ANN and could detect OSA stages with high accuracy (92.3%).³⁹ Accuracies of 97.1% for per-recording classification and 86.2% for per-segment OSA detection with satisfactory sensitivity and specificity were achieved using a discriminative hidden Markov Model from ECG Signals.⁴⁰ An accuracy of 92.85% was demonstrated using SVMs for automated recognition of OSA using nocturnal ECG recordings.⁴¹ A CNN using 1-lead ECG signal showed a good OSA detection performance with an average accuracy, recall, and specificity of 92.4%, 92.3%, and 92.6%.⁴² Using single-lead ECG signals, a hybrid “K-means, Recursive least squares” radial basis function network showed a suitable OSA detection percentage near 96%.⁴³ Furthermore, ECG, SpO₂, and body mass index used as inputs in a multilayer feedforward neural network, achieved better performance than other ML methods.⁴⁴ Using single-lead ECG data, a CNN showed 96% precision, recall, and F1-score to detect sleep apnea events.⁴⁵

In the setting of sleep apnea and hypopnea syndrome (SAHS), using the spectral HRV as the input signal has been used with a good performance for screening purposes.⁴⁶ Sharma et al⁴⁷ applied the least-square SVM classifier with the Gaussian radial basis function kernel on a single-lead ECG and identified apnea and hypopnea events on the minute-by-minute basis with an accuracy of about 84%. Furthermore, an ANN using ECG signals showed an accuracy of 90% to classify apnea, hypopnea, and normal breathing events.⁴⁸

Electroencephalogram-based diagnosis

Electroencephalogram (EEG) signals were used as input features in an SVM and showed a 95.1% accuracy in detecting apnea patients.⁴⁹ Another study that applied SVM, kNN, and ANN models on EEG signals to detect sleep apnea showed better accuracy using SVM (99%) compared to kNN (75%) and ANN (86%).⁵⁰ ANN has also been used to detect OSA using pupil size and EEG⁵¹ and showed a high accuracy rate in differentiating OSA and narcolepsy.

SpO₂-based diagnosis

Application of a novel ANN on nocturnal pulse oximetry with demographic, anatomic, and clinical data to detect sleep-disordered breathing,⁵² showed a good performance with area under curve (AUC) varying between 0.904 and 0.954 depending on the apnea-hypopnea index (AHI) threshold.⁵² In patients with a clinically suspected OSA, approximate entropy analysis of arterial oxygen saturation data showed higher sensitivity and specificity in diagnosing OSA than traditional methods.⁵³ At home OSA diagnosis using oximetry and airflow signals showed better performance than single-channel approaches,⁵⁴ highlighting the importance of ML techniques in the diagnosis of OSA at home. Overnight oximetry using SVM models has been found to accurately diagnose severe OSA.⁵⁵ In another study, a multilayer perceptron ANN using portable oximetry recording was examined as a screening test for moderate-to-severe OSA diagnosis.⁵⁶ The authors found a good performance (92.4% sensitivity and 95.9% specificity) of the proposed model both in patients with nonchronic obstructive pulmonary disease and those with chronic obstructive pulmonary disease. Overnight oxygen saturation and a probabilistic NN have been found to outperform the existing univariate and multivariate models in OSA detection.⁵⁷ A multilayer perceptron neural network using nocturnal SpO₂ recordings provided a diagnostic accuracy of 85.5% (89.8% sensitivity and 79.4% specificity) in detecting OSA.⁵⁸ Similarly, a CNN applied on nocturnal SpO₂ measured in patients with stroke or transient ischemic attack, detected moderate-to-severe sleep apnea with a sensitivity of 92.3% and specificity of 96.1%, respectively.⁵⁹ Interestingly, in a nonreferred population, oximetry data combined with demographic characteristics has been used effectively as a screening tool to identify OSA.⁶⁰ An SVM has been used for the automatic detection of apnea/hypopnea events using SpO₂ signals achieving greater than 90% accuracy.⁶¹ SpO₂ and ECG features used as inputs in a classifier combination of AdaBoost with Decision Stump, Bagging with REPTree, and either kNN or Decision Table achieved an accuracy of 82% for a minute-based real-time SAHS detection.⁶²

Sound-based diagnosis

Respiratory sounds have also been used for the detection of OSA using a CNN.⁶³ The proposed method outperformed other models and achieved an accuracy of 80.17% for AHI threshold 15 and 80 events/h, 21% for AHI threshold 80.21 events/h for all patients.⁶³ By including only males, the accuracy of the proposed method was 81.63% for AHI threshold 15 events/h and 77.22% for AHI threshold 30 events/h.⁶³ CNN and recurrent NN have been studied in snore detection, and the proposed algorithm was found to achieve an accuracy of 95.3% over all microphones in snore detection.⁶⁴ A snoring detection method using a polyvinylidene fluoride sensor has also been proposed.⁶⁵ An SVM used in snore detection showed a sensitivity and positive predictive value of 94.6% and 97.5%, respectively.⁶⁵ Tiron et al⁶⁶ showed that a novel hybrid acoustic smartphone app technology performed better compared to the gold standard PSG in detecting clinically significant OSA, as well as in estimating AHI. Specifically, the performance of the proposed technology had a sensitivity of 88.3% and specificity of 80.0% for a clinical threshold for the AHI of ≥ 15 events/h of detected sleep time.⁶⁶

Application of logistic regression on acoustic features from overnight snoring sounds identified SAHS,⁶⁷ while a long short-term memory model has also been found to achieve the highest accuracy compared to CNN, using snore signals in detecting SAHS.⁶⁸ Tracheal sounds used as input features in a deep neural network to detect apnea events showed an AUC of 0.99 and 0.98 for diagnosis of sleep-disordered breathing using an AHI cutoff of > 5 and>30 events/h, respectively.⁶⁹

Other signals used for diagnostic purposes

An ML model that included a dual accelerometry system for airflow estimation showed an accuracy of 89% in estimating AHI.⁷⁰ An RF model using a wearable respiratory effort signal with or without SpO₂ signal has been shown to accurately predict AHI, showing the best performance when using both signals.⁷¹ Automated detection of sleep apnea events can be effectively achieved using a deep bidirectional long short-term memory-based detection model⁷² using the nasal pressure signal.⁷² In a pilot study, application of an SVM learning algorithm lead to the distinction between normal and apnea waveforms, with a classification accuracy of 97.22% for power ratio and reflection index.⁷³ The SVM and the RF classifiers were used for the automated detection of sleep apnea using bivariate cardiopulmonary signals, and the approach demonstrated an average sensitivity and specificity of 82.27% and 78.67%, respectively, in detecting sleep apnea.⁷⁴ ML algorithms and especially a deep learning CNN framework outperformed existing methods in identifying OSA in a specific population of patients with stroke.⁷⁵ The role of multilayer-perceptron NN as an assistant tool in diagnosing OSA has also been studied,⁷⁶ and the proposed algorithm showed a good performance in this setting in reducing the demand for OSA screening in polysomnographic studies.⁷⁶ A fuzzy NN using easily obtained variables has also been proposed as a tool for the prediction of moderate to severe obstructive apnea-hypopnea syndrome.⁷⁷ Interestingly, a real-time diagnosis of sleep apnea using an SVM classifier and a smartphone-based analysis has shown high accuracy of OSA diagnosis.⁷⁸

Also, ML techniques have been used to assess the correlation between snoring and the presence and severity of apnea,⁷⁹ exhibiting that the snore index had a weak negative correlation with central AHI and modest positive and negative predictive values for OSA.⁷⁹ Multivariable nonintrusive measurements have been used to screen for OSA.⁸⁰ Specifically, a multifeature vector has been employed by a NN for OSA/non-OSA classification showing good (sensitivity of 91% and a specificity of 89% for test data) performance.⁸⁰

Erdenebayar et al⁸¹ proposed a novel method for OSA detection using a piezoelectric sensor. An SVM was used as a classifier to detect OSA events. The method achieved a mean accuracy of 71.5%, 80.0%, and 71.9% for the mild, moderate, and severe groups, respectively.⁸¹ Stage-one decision tree and stage-two linear regression models were used to analyze signals from pressure sensors installed in patients’ home bed,⁸² showing a detection rate of 82.9% and an OSA severity classification accuracy of 74.3%.⁸² An SVM using polynomial kernel showed better performance and the highest accuracy of 82.4% (sensitivity: 69.9%, specificity: 91.4%) for OSA detection using HRV, SpO₂, and the respiratory effort signals.⁸³ An SVM with predicted aerodynamic and biometric features showed a classification accuracy, sensitivity, specificity, and F1-score for moderate OSA diagnosis of 81.5%, 89.3%, 86.2%, and 87.6%, respectively.⁸⁴ ML models have been successfully used for respiratory scoring in patients with sleep apnea using PPG and HRV signals.⁸⁵

Furthermore, the SVM and the kNN algorithms have been used to classify patients and healthy individuals, exhibiting a good performance in diagnosing OSA.⁸⁶ Using clinical variables, multilayer perceptron classifiers resulted in diagnostic accuracy of 86.6%.⁸⁷ Yue et al⁸⁸ proposed a system based on residual networks and single-channel nasal pressure airflow signals, which showed a correlation for AHI between OSA smart system and the registered polysomnographic technologist score (determined by 2 technologists) to be 0.94 (P < .001) and 0.96 (P < .001), respectively. An SVM based model using facio-cervical measurements showed a good performance in the screening of moderate to severe OSA, especially for asymptomatic patients.⁸⁹ Beyond the high accuracy of sleep apnea diagnosis, a reduction of the cost and time in the diagnosis of OSA has been obtained using a visual-scoring based algorithm and a morphological filter via ANN using respiratory signals.⁹⁰ ML algorithms based on ballistocardiogram data have been also proposed as an effective method for the detection of OSA.⁹¹ A deep CNN that employed data from lateral cephalometric radiographs for image-based detection of OSA exhibited an AUC of 0.92.⁹² A wrist‐worn reflective photoplethysmography (PPG) employing a deep learning model showed a sensitivity and specificity of 77% and 75%, respectively (AHI threshold 10 events/h) in estimating AHI.⁹³ Finally, oronasal airflow, along with the thoracic and abdominal respiratory effort signals, were analyzed using long short-term memory and showed a good performance (overall accuracies in the databases: 80.66%/82.04%) in the discrimination of apnea/hypopnea events.⁹⁴

Detection of sleep apnea in children

ML algorithms have been used in the pediatric population to identify those patients at risk of obstructive apnea/hypopnea syndrome. Specifically, Calderon et al⁹⁵ showed that an LR algorithm outperformed the SVM and the AdaBoost algorithm in this setting. A multilayer perceptron NN with a Bayesian approach using airflow signal by means of recurrence plots features showed a good performance in diagnosing pediatric sleep apnea.⁹⁶ In children with clinical symptoms suggestive of OSA syndrome, overnight oximetry processed via Bluetooth technology by a cloud-based machine learning-derived algorithm was found to diagnose obstructive SAHS with good performance (accuracy > 79% for AHI estimates of 1–10 events/h).⁹⁷ In sleeping infants, uncalibrated signals of thoracic and abdominal respiratory movements that were analyzed using ANN aiming to detect obstructive apnea events were found to be insufficient in the detection of apnea events in infants.⁹⁸

Respiratory activity assessed by monitoring sleep mandibular movement has been used for the detection of moderate to severe OSA in the pediatric population.⁹⁹ Furthermore, a multilayer perceptron model that was used to estimate AHI severity in children showed high accuracy in diagnosing moderate (AHI = 5 events/h) and severe (AHI = 10 events/h) SAHS: 81.3% and 85.3%, respectively.¹⁰⁰ Finally, in a pediatric population, an ML-based screening tool was used to identify children needing overnight monitoring for OSA following tonsillectomy or adenoidectomy¹⁰¹; combining oximetry and actigraphy, this model showed an accuracy of 87–89% for AHI > 2 events/h and 95–96% for AHI > 10 events/h.¹⁰¹

Prediction of sleep apnea

Taghizadegan et al¹⁰² proposed an ensemble of recurrence plots and pretrained convolutional neural networks to predict OSA using single EEG and ECG signals, outperforming the state-of-the-art methods. A boosting algorithm has shown a good performance (AUC: 0.747) in predicting mild obstructive sleep-disordered breathing.¹⁰³ On the other hand, an explainable fuzzy NN, a back propagation NN, and a stepwise regression model that employed easily obtained variables including waist circumference, mean blood pressure at the end of polysomnography, and the difference in systolic blood pressure between the end and start of PSG indicated that, although the explainable fuzzy NN should be the preferred method to predict moderate-to-severe OSA, none of the tested methods had good efficacy in predicting the AHI values.⁷⁷

Supervised ML methods for prediction purposes were proposed by Keshavarz et al.¹⁰⁴ The results showed that the best prediction model was the naive Bayes and LR classifiers (AUC 0.768 and 0.761, respectively). At the same time, the SVM and the naive Bayes could be used for screening high-risk people with OSA. Schwartz et al¹⁰⁵ proposed the ElasticNet algorithm for predicting OSA. The algorithm showed a good performance, with an overall sensitivity of 83.4% and a specificity of 66.5%, while the overall accuracy of the model was 73.2%.¹⁰⁵ In a population from South Korea, in models that used logistic regression, SVM, RF, and XGBoost to predict OSA,¹⁰⁶ the SVM showed the best (AUC: 0.87) and XGB showed the lowest (AUC: 0.80) prediction performance, and the XGB showed the lowest OSA.¹⁰⁶

Clinical variables were also used as input features in a generalized regression NN to predict OSA. The trained model showed an accuracy of 91.3% with a sensitivity and specificity of 98.9% and 80%, respectively.¹⁰⁷ Mihaicuta et al¹⁰⁸ proposed an algorithm that consisted of a classification tree and a computational framework and showed a significant specificity improvement for only an 8.2% sensitivity decrease compared to the state-of-the-art STOP-BANG. Another SVM model that was constructed to predict OSA based on 3 anthropometric features (neck circumference, waist circumference, and body mass index) and age,¹⁰⁹ showed that dividing by sex and age for the AHI threshold 15 events/h, the cross-validation and testing accuracies in young females were 85.3% and 76.7%, respectively.¹⁰⁹

Also, ML-based models aiming to predict OSA in patients with Down syndrome have been proposed. Specifically, using a logic learning machine, the best model had a negative predictive value of 73% for mild OSA and 90% for moderate or severe OSA¹¹⁰; on the other hand, the positive predictive values were 55% and 25%, respectively.¹¹⁰

El-Solh et al¹¹¹ proposed an ANN using anthropomorphic measurements and clinical information to predict the AHI, showing that ANN could be a valuable tool to predict OSA. Ustun et al¹¹² proposed an ML method known as Supersparse Linear Integer Models and showed that age, sex, body mass index, and medical history were superior to the symptom variables in predicting OSA. Teferra et al¹¹³ proposed OSUNet, a prediction tool for OSA based on ANN, and showed that in the validation group, STOP-BANG, and modified neck circumference (MNC) had higher sensitivities compared with the OSUNet, but the STOP-BANG and the MNC had lower specificities compared with the OSUNet. Furthermore, the OSUNet had the highest positive predictive value.¹¹³ Huang et al¹¹⁴ proposed an SVM model to predict AHI scores using clinical variables. The SVM model performed better exhibiting a more balanced sensitivity and specificity than the LR, the Berlin questionnaire, the NoSAS Score, and the Supersparse Linear Integer Model scoring system. Demographic characteristics, spirometry values, gas exchange (PaO₂, PaCO₂), and symptoms have been used as input features to predict the severity of AHI.¹¹⁵ Using these features, SVM and linear regression showed to better predict AHI.¹¹⁵ Kim et al¹¹⁶ examined various ML models using data from breathing sounds obtained by a noncontact device to predict AHI. Although the examined models depicted a similar performance, RF resulted in the highest performance compared to Gaussian process, SVM, and simple linear regression.¹¹⁶

Ferre et al¹¹⁷ used multiple LR and the unbiased recursive partitioning technique conditional inference tree to detect patients at high risk of sleep-related breathing disorders; it was shown that although both models can be used for the prediction of sleep-related breathing disorders, the unbiased recursive partitioning technique conditional inference tree had a higher specificity and was easier to be implemented in clinical practice.¹¹⁷

Finally, ANN has been investigated for the prediction of apnea and hypopnea with a sensitivity and specificity of up to 80.6% and 72.8%, respectively¹¹⁸; hypopnea prediction achieved a sensitivity and specificity of 74.4% and 68.8%, respectively.¹¹⁸

Sleep apnea classification

ML algorithms have also been used to classify apnea into central, obstructive, or mixed events. In this setting, a feedforward NN showed an accuracy of approximately 84%.¹¹⁹ Furthermore, an ANN that used abdominal effort signals has been used for the classification of sleep apnea into obstructive, central and mixed, showed an accuracy of 85.62% in classifying sleep apnea syndrome using this method.¹²⁰ Biswal et al¹²¹ showed that a deep recurrent and CNN model achieved an overall diagnostic accuracy of 88.2% in converting AHI values into standard clinical categories of mild, moderate, and severe disease. Baty et al¹²² showed that the classification of OSA severity is feasible using signals from an ECG belt; interestingly, the signals from the ECG belt were comparable to the patched ECG.

Another model integrating fuzzy set theory and decision tree based on anthropometric and questionnaire data outperformed other methods (LR, decision tree, backpropagation neural network, SVM, and learning vector quantization) regarding the prediction of the OSA severity.¹²³ A CNN using nocturnal lead ECG signals showed an accuracy of 99.0% in identifying mild and moderate sleep apnea.¹²⁴ A CNN using ECG signals has been proposed for multiclass classification of OSAH exhibiting an F1-score of 87.0% for the test set discriminating normal, apnea, and hypopnea events.¹²⁵ An SVM showed an accuracy of 92.3% in classifying patients as moderate or severe OSA compared to healthy individuals using as inputs audio, actigraphy, PPG, and demographics.¹²⁶ An SVM using ECG signals showed a specificity, sensitivity, and subject-based classification accuracy of 95.20%, 92.65%, and 94.3%, respectively, in classifying patients into normal or apnea groups.¹²⁷ Another SVM using PPG and SpO₂ signals showed sensitivity and positivity predictive values of 74.2% and 87.5% for apnea, 87.5% and 63.4% for hypopnea, and 92.4% and 92.8% for apnea + hypopnea, respectively.¹²⁸ A linear SVM using speech signals showed that the accuracy of AHI classifications using 2 AHI thresholds, 30 and 10 events/h, were both 78.8%, the sensitivities were 77.3% and 79.1%, and the specificities were 80.3% and 78.0%, respectively.¹²⁹ A combination of convolutional and recurrent NN has been used for OSA severity classification based on PPG signals, especially sleep continuity.¹³⁰

A deep learning framework has been used to classify patients into healthy patients, patients with OSA, patients with restless legs syndrome, and patients with both OSA and restless legs syndrome. The proposed model achieved a mean accuracy of 72% and a weighted F1 score of 0.57.¹³¹ Long short-term memory using a PPG signal has been used for multiclass classification of sleep apnea, including normal, apnea, and hypopnea. The positive predictive value was 94.16% for normal, 81.38% for apnea, and 97.92% for hypopnea.¹³² Simple logistics, SVM, and deep NN, have been applied for severity classification using breathing sounds; 10-fold cross-validation exhibited an accuracy of 88.3% in the 4-group classification and an accuracy of 92.5% in the binary classification.¹³³ An ANN using SpO₂ signals recorded during ambulatory polygraphy showed an accuracy of 90.9% in the estimation of the AHI.¹³⁴ A Long Short-Term Memory NN using input signals from peripheral blood oxygen saturation, thermistor-airflow, nasal pressure-airflow, and thorax respiratory effort showed an overall accuracy of 87% in classifying OSA into the standard severity groups.¹³⁵ A one-dimensional CNN achieved a mean F1 score of 92% for normal breathing, 87% for central sleep apnea, 72% for coughing, 51% for obstructive sleep apnea, 57% for sighing, and 63% for yawning using accelerometer and gyroscopic data.¹³⁶ A feedforward NN was used to discriminate apnea and hypopnea events from normal breathing events using ECG signals as inputs, showed an accuracy for the apnea and hypopnea detection of 94.72% and 79.77%, respectively.¹³⁷ RF models have been found to outperform the regularized logistic regression models in classifying sleep apnea patients using tracheal breathing sounds.¹³⁸ A long short-term memory recurrent NN has also been evaluated in classifying sleeping breathing patterns into OSA, central sleep apnea, hypopnea events, and normal breathing with an overall accuracy of 92.3%.¹³⁹

Therapy optimization and other applications of ML in sleep apnea management

Optimization of the continuous positive airway pressure (CPAP) therapy is a time-consuming process and may require sleep laboratory stay(s) for CPAP titration. In recent years the use of auto-titrating PAP devices without a prior manual titration has been recommended.¹⁴⁰ Use of an ANN algorithm has successfully predicted optimal CPAP levels.¹⁴¹ ML models have also been used to optimize CPAP titration. Specifically, an ANN-guided CPAP titration arm was shown to achieve optimal CPAP pressures in a shorter time interval compared to the control group,¹⁴² which may reduce CPAP titration failure and obviate the need for repeat in-laboratory titration for some patients. Prediction of compliance with CPAP therapy is another field of ML implementation,¹⁴³ with important implications for helping clinicians detect individuals with OSA who may struggle and need more support with initiation of CPAP therapy.

Another area of ML implementation is the identification of patients who will respond to oral appliance therapy. In this setting, an unattended, in-home feedback-controlled mandibular positioner test showed a good performance in predicting patients who will respond to oral appliance therapy (positive predictive value 97%; and negative predictive value 72%).¹⁴⁴ The prediction of surgical success in patients with OSA is of great clinical importance as well. For surgical success prediction, the gradient boosting model showed the best performance, with a reported accuracy of 70.8%.¹⁴⁵ In another study involving the prediction of outcomes in palatal surgery, Lasso exhibited the best performance among the other predictive models (Elastic Net, Ridge, Bagging, RF, SVM, NN).¹⁴⁶ Furthermore, multiclass linear discriminant analysis has been used to predict the site of upper airway collapse in OSA.^147,148 which may help clinicians and their patients choose the most appropriate and individualized treatment option.

Other applications of ML models have also been proposed. Specifically, a Fuzzy Analytic Hierarchy Process approach showed that nighttime symptoms could be used to prioritize polysomnography queueing,¹⁴⁹ which would allow clinicians to expedite the evaluation of individuals most likely to have severe OSA. Furthermore, a deep NN has been used to predict post-apnea systolic and diastolic blood pressure using clinical PSG signals.¹⁵⁰ Blood pressure prediction can help clinicians to predict cardiovascular risk.¹⁵⁰

ML models can also be implemented for the selection of the best OSA diagnostic tool. Specifically, ML models have been found to outperform conventional methods for the prediction of patients with nondiagnostic home sleep apnea tests and, therefore will require in-laboratory PSG.¹⁵¹ LArge Memory STorage and Retrieval ANN showed a sensitivity of 81% and specificity of 77% in predicting OSA-related disordered breathing events.¹⁵²

Quality of evidence and risk of bias

The quality assessment of the 132 studies we included met the criteria of QUADAS-2. Detailed results of risk of bias are shown in the supplemental material. Among these, high risk of bias of patient selection is notable, which may underline the limitation of these studies with respect to the heterogeneity of the patient population. The risk of bias and concern of applicability was shown in Figure S1 ^{(243.7KB, pdf)} in the supplemental material.

DISCUSSION

This review shows that the ML techniques can play a crucial role in diagnosis and management of sleep apnea. Currently, the diagnosis of sleep apnea is time-consuming and cumbersome process, often requiring an overnight stay in the sleep laboratory. PSG remains the gold standard method for the diagnosis and severity of sleep apnea, during which clinicians can obtain important information pertinent to a patient’s underlying physiological state. However, the main limitation of this approach is the disruption of the patients’ sleep caused by the multiple sensors and the hospital environment. Although home sleep apnea tests are widely utilized, the accuracy of the diagnosis or severity estimation of OSA with these devices is reduced.¹⁵³

Overall, ML-based diagnosis of sleep apnea is feasible and has demonstrated good performance; furthermore, many wearable technologies have been introduced to address the sleep apnea diagnosis.^154,155 Different input features have been used, including ECG data, SpO₂ signals, respiratory signals, EEG data, sound data, and data from pressure sensors.¹⁵⁶ In some models, clinical information is incorporated as well. Huo et al¹⁵⁷ developed a ML-based questionnaire consisting of 2 logistic regression classifiers using clinical information from 2 large observational cohort studies. The model outperformed three commonly used OSA screening questionnaires (4 variable, STOP-BANG and Berlin). ML can be used to predict adverse outcomes associated with OSA. Recently, Li et al¹⁵⁸ demonstrated that in a large cohort of individuals with OSA, RF modeling using AHI, clinical, anthropometric, and demographic information was able to predict 10-year cardiovascular disease mortality.

Another interesting area of ML implementation in the field of sleep apnea is the prediction of sleep apnea severity (AHI). In this setting, ML models have used clinical data, anthropometric features, sound signals, and ECG. Furthermore, ML has been used for classification purposes, to distinguish obstructive, central, and mixed apnea events. Photoplethysmography, ECG, SpO₂, and sound signals were mainly used as inputs in this setting. ML models have also been implemented for the prediction of outcomes following surgical techniques, assessment of compliance with CPAP therapy, and determining optimal CPAP pressure.

In general, prospective studies are needed not only to further establish the accuracy and generalizability of these algorithms outside their development cohorts irrespective of their size, but also their translation to actionable care pathways that can demonstrate clinical utility. Generalizability issues pertinent to ML algorithms may arise from differences between organizations and regions or results and outcomes that may vary in time, such that they do not match those of the original data source.¹⁵⁹ It is thus imperative that AI developers understand and communicate their algorithm-design choices and study assumptions with clinicians. Also, researchers should adopt processes that help avoid introducing bias. A causal diagram can be helpful to infer the generalizability of models, by making explicit which relationships in the data are likely to differ between organizations and/or across time.¹⁶⁰ Finally, model evaluation should be tailored to the intended use of the system.¹⁶¹ These processes must be organically integrated to avoid bias and maximize generalizability of findings, therefore avoid perpetuating existing healthcare inequalities.

Furthermore, the efficacy of ML algorithms should be Food and Drug Administration “labeled” with respect to the subject populations that have been evaluated. As new patient groups are studied, thereby reducing the sample bias, such descriptions should be entered into the ML algorithm label.¹⁶²

Although existing data show promising results, significant obstacles need to be overcome for the implementation of ML models for the diagnosis and management of sleep apnea in the real-world setting. In general, similarly with other domains in medicine, a framework for implementing ML algorithms in clinical practice is needed.¹⁶² For example, should they be used only in the context of screening high risk populations for sleep apnea? Alternatively, are some algorithms sufficiently accurate that they can be used as the definitive test to guide diagnosis and treatment? Some of the main challenges that also should be addressed include the physician liability, data protection, protection of patient’s rights, system failure reporting, system upgrading, and cybersecurity issues.^162,163

Limitations

Some limitations should be reported. Most of the included studies consisted of small sample size, and the results should be interpreted with caution. The acquisition of input features should be of high quality, and the absence of artifacts is necessary. This is a limitation of ML techniques that restrain their use in clinical practice. Furthermore, some studies included patients who were referred for PSG, and as a result, they had a high pretest probability of the disease. Therefore, the results of each included study should be interpreted in the light of inclusion criteria. Averaging time and sampling frequency of the pulse oximeter can alter the oxygen desaturation index. These parameters can differ between studies and therefore they should be taken into consideration in the interpretation of the results. Another limitation of these studies is that should be noted is that no ML technique has been proposed for the detection of apneas without an associated oxygen drop. Therefore, this type of apnea is currently underdiagnosed using the existing algorithms.

CONCLUSIONS

The implementation of ML in the field of sleep apnea appears to be rapidly growing and promising. Advancements in wearable sensor technology, signal processing, and ML models can help clinicians predict, diagnose, and classify sleep apnea more accurately. Also, ML models can provide essential data to optimize the treatment strategy in patients with sleep-related breathing disorders.

DISCLOSURE STATEMENT

All authors have seen and approved this manuscript. The work was supported by the Institute of Precision Medicine (17UNPG33840017) from the American Heart Association; the RICBAC Foundation; and National Institutes of Health Grants 1 R01 HL135335-01, 1 R21 HL137870-01, 1 R21EB026164-01, 3R21EB026164-02S1, and 1 R01 HL161008-01. Dr. Quan was partially supported by National Institutes of Health Grant R21 HL159661. The authors report no conflicts of interest.

ABBREVIATIONS

AHI

apnea-hypopnea index

ANN

artificial neural network

AUC

area under curve

CNN

convolutional neural network

CPAP

continuous positive airway pressure

ECG

electrocardiogram

EEG

electroencephalogram

HRV

heart rate variability

kNN

k-nearest neighbor

LR

linear regression

ML

machine learning

OSA

obstructive sleep apnea

PPG

photoplethysomography

PSG

polysomnography

RF

random forest

SAHS

sleep apnea and hypopnea syndrome

SpO₂

oxygen saturation

SVM

support vector machine