A Novel Deep Learning Approach for Diagnosing Sleep Apnea Using Feature Fusion of ECG and SpO2 Signals

Abstract

Introduction: Frequent disruptions in breathing during sleep also known as Sleep Apnea (SA), is a common sleep disorder, that poses serious health concerns all across the world. Global prevalence of SA is very high, around 936 million adults are suffering from this disorder worldwide. Primary causes of SA include Obesity, old age, being male, high BMI and some other causes are smoking, alcohol, opium consumption etc. If untreated on time, has severe consequences like morning head ache, daytime sleepiness, fatigue, hypertension, diabetes, cognitive impairments and in some cases, it extends to cardiovascular diseases, stokes as well.
Objectives: The aim of this research-study is to identify sleep apnea events through the analysis of Electrocardiogram (ECG / EKG), Blood Oxygen saturation level (SpO2) signals.
Methods: The study employs the PhysioNet Apnea ECG 1.0.0 dataset for training a machine learning/deep learning algorithm. The proposed system processes ECG and SpO2 data concurrently, with machine learning models trained individually for each type of signal. ECG signals offer crucial insights into heart rate variability and arrhythmias, while SpO2 measurements reveal variations in blood oxygenation during sleep. Training models on these individual signals allows for the capture of unique properties significant to sleep apnea identification. A new feature space is formed by concatenating the features extracted from both these signals and then a 1D-CNN model was trained-tested on this new feature set, enhancing the overall accuracy of predictions. Using ECG and SpO2 data, this model accurately identifies apnea occurrences.
Results: The technique yielded promising results, potentially enhancing the early-stage diagnosis and treatment recommendation for sleep-apnea. Our research analysis attained Accuracy, Specificity and AUC of 91%, 92% and 0.93 respectively.
Conclusions: Using multimodal approach like ECG and SpO2, performance of Sleep-apnea predicting models can be increased to a level that physicians can rely on. Future research will explore the integration of additional physiological signals like limb movement, chest and abdomen movement etc. and generate recommendations for sleep apnea patients by building recommender systems on top of these results.