Detection of sleep apnea using Machine learning algorithms based on ECG Signals: A comprehensive systematic review

• Published in: Expert systems with applications Vol. 187; p. 115950
• Main Authors: Salari, Nader, Hosseinian-Far, Amin, Mohammadi, Masoud, Ghasemi, Hooman, Khazaie, Habibolah, Daneshkhah, Alireza, Ahmadi, Arash
• Format: Journal Article
• Language: English
• Published: New York Elsevier Ltd 01.01.2022; Elsevier BV
• Subjects: Accuracy; Algorithms; Artificial neural networks; Classification; Deep learning; Diagnosis; Electrocardiogram; Machine Learning; Neural networks; Polysomnography; Segments; Sensitivity; Sleep Apnea; Support vector machines; Systematic Review; Accuracy; Polysomnography; Machine Learning; Electrocardiogram; Systematic Review; Sleep Apnea
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2021.115950

•Diagnosis of sleep apnea based on ECG characteristics is very accurate.•Diagnosis of sleep apnea with electrocardiogram can replace the present methods.•SVM and Neural Network algorithms were highly accurate.•Frequency and time domain features were the most commonly used features. Sleep apnea (SA) is a common sleep disorder that is not easy to detect. Recent studies have highlighted ECG analysis as an effective method of diagnosing SA. Because the changes caused by SA on the ECG are imperceptible, the need for new methods in diagnosing this disease is required more than ever. Machine Learning (ML) is recognized as one of the most successful methods of computer aided diagnosis. ML uses new methods to diagnose diseases using past clinical results. The purpose of this study is to evaluate studies using ML algorithms based on ECG characteristics to assess people suffering from SA. In this study, systematically-reviewed articles written in English before October 2020 and indexed in PubMed, Scopus, Web of Science, and IEEE databases were searched with no lower time limit. From these articles, 48 were selected for further review. The selected articles adopteddifferent ML methods for classification. All of these studies were binary where SA was detected from the normal state based on a full ECG stripe (per record), or based on one-minute segments (per segment). Our analysis show that the most common features used in the studies were frequency, time series, and statistical features. Support-Vector Machine (SVM) and deep learning-based neural network (i.e. CNN, DNN) performed best in full record data detection. The highest accuracy, sensitivity, and specificity reported among the selected studies were 100%, which was obtained by an SVM. In another study, the classification was conducted based on ECG segments, and accordingly, the highest classification accuracy was observed in the residual neural network algorithm (RNN). The accuracy, sensitivity, and specificity of this algorithm were reported to be 99%. In general, it can be stated that ML techniques based on ECG characteristics have a high capability in diagnosing SA. These techniques can increase the diagnosis of patients with SA or the detection of SA episodes on ECG record, and can potentially prevent complications of the disease at later stages.