Accurate detection of sleep apnea with long short-term memory network based on RR interval signals

• Published in: Knowledge-based systems Vol. 212; p. 106591
• Main Authors: Faust, Oliver, Barika, Ragab, Shenfield, Alex, Ciaccio, Edward J., Acharya, U. Rajendra
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 05.01.2021; Elsevier Science Ltd
• Subjects: Accuracy; Algorithms; Bandpass filters; Deep learning; Detrending; Diagnosis; Electrocardiography; Gaussian process; Heart rate variability; Intervals; Machine learning; Sensitivity; Signal classification; Signal processing; Sleep apnea; Deep learning; Detrending; Sleep apnea; Heart rate variability
• ISSN: 0950-7051; 1872-7409
• DOI: 10.1016/j.knosys.2020.106591

Sleep apnea is a common condition that is characterized by sleep-disordered breathing. Worldwide the number of apnea cases has increased and there has been a growing number of patients suffering from apnea complications. Unfortunately, many cases remain undetected, because expensive and inconvenient examination methods are formidable barriers with regard to the diagnostics. Furthermore, treatment monitoring depends on the same methods which also underpin the initial diagnosis; hence issues related to the examination methods cause difficulties with managing sleep apnea as well. Computer-Aided Diagnosis (CAD) systems could be a tool to increase the efficiency and efficacy of diagnosis. To investigate this hypothesis, we designed a deep learning model that classifies beat-to-beat interval traces, medically known as RR intervals, into apnea versus non-apnea. The RR intervals were extracted from Electrocardiogram (ECG) signals contained in the Apnea-ECG benchmark Database. Before feeding the RR intervals to the classification algorithm, the signal was band-pass filtered with an Ornstein–Uhlenbeck third-order Gaussian process. 10-fold cross-validation indicated that the Long Short-Term Memory (LSTM) network has 99.80% accuracy, 99.85% sensitivity, and 99.73% specificity. With hold-out validation, the same network achieved 81.30% accuracy, 59.90% sensitivity, and 91.75% specificity. During the design, we learned that the band-pass filter improved classification accuracy by over 20%. The increased performance resulted from the fact that neural activation functions can process a DC free signal more efficiently. The result is likely transferable to the design of other RR interval based CAD systems, where the filter can help to improve classification performance.