Comparing the performances of deep learning model with different signals underlying resampling techniques to classify sleep apnea

Abstract

Millions of people around the world are suffering from long term Sleep Apnea. Full scale sleep test is costly and time-consuming. This paper, using deep learning model, chose a single candidate signal from multi-channel polysomnogram data for sleep apnea screening. Nature of data, however, shows an imbalance class of dataset between normal and apneic events. To increase the binary classification output performance, two resampling techniques; Synthetic Minority Over-sampling Technique (SMOTE) and Random Under-Sampling (RUS), were employed in Bidirectional Long Short-Term Memory (Bi-LSTM) model training. One hundred polysomnography (PSG) records were randomly selected from the Multi-Ethic of Atherosclerosis (MESA) database in this study. They were trained under three conditions; original, SMOTE and RUS datasets. Our results showed (1) Cohen's kappa score was greater in resampling (SMOTE, RUS) datasets than original one. (2) Between the resampling techniques, metrices in SMOTE were better than ones in RUS. (3) Within SMOTE, the abdominal belt was the best among other signals with Cohen's kappa score of 0.2078 and 58.99% in F1-score. These findings suggested that abdominal belt was the best candidate signal for sleep apnea screening.