407 Explanatory analysis of polysomnography for the identification of sleep apnea hypopnea events using deep learning neural network

Abstract

Abstract Introduction Using deep learning algorithms, we investigated univariate and multivariate effects of four polysomnography features including heart rate (HR), electrocardiogram (ECG), oxygen saturation (SpO2) and nasal air flow (NAF) on the identification of sleep apnea and hypopnea events. This explanatory analysis that may clarify the sensitivity and specificity of those features to SAs and SHs have not been probed. Methods We studied 804 polysomonography samples from 704 patients with obstructive sleep apnea and 100 controls. The input data were converted into scalograms as 4-channel 2D images to train Xception networks. For training, 77,638 patches were sampled from the original 6-hour sleep data with 30-second time width. A 10% of these patches were segregated as the test-set. With each feature sets, we tested the following classifications: 1) normal vs apnea vs hypopnea; 2) normal vs. apnea+hypopnea; 3) normal vs. apnea; and 4) normal vs. hypopnea. Results SpO2 classified normal vs. apnea most accurately (98%), followed by NAF (85%), ECG (77%), and HR (63%). SpO2 also showed the highest accuracy in classifying normal vs. hypopnea (87%), and normal vs. apnea+hypopnea (96%) and three groups (82%). When the combination of four features were used, the classification accuracies were generally improved compared to use of SpO2 only (normal vs. apnea 99%; vs. hypopnea 89%; vs. apnea+hypopnea: 94%; three groups: 86%). Conclusion Deep learning with SpO2 or NAF feature most accurately classified apneas from normal sleep events, suggesting these features’ characterization of sleep apnea events. Oxygen desaturation, which is a typical pattern of hypopnea, was only the feature showing reliable accuracy in classifying hypopnea vs. normal. Nevertheless, combination of four polysomnography features could improve the identification of sleep apnea and hypopnea. Furthermore, classifying normal vs. apnea+hypopnea was more accurate than separately classifying three groups, suggesting deep learning approaches as the primary screen tool. Since the classification accuracy of using SpO2 was higher than any other features, developing a portable equipment measuring SpO2 and running deep learning algorithms has the potential for inexpensive, accurate diagnostics of obstructive sleep apnea syndrome. Support (if any) This study was supported by USC STEVENS CENTER FOR INNOVATION TECHNOLOGY ADVANCEMENT GRANTS (TAG), BrightFocus Foundation Award (A2019052S).