Effect of Class Weights on Imbalanced Classes in Bi-directional LSTM Training for Sleep Apnea Classification

Sleep apnea, which is defined as the repetitive cessations of breathing during sleep, is the common disorder worldwide. The cost and the process of sleep test to obtain the polysomnogram is not optimal for sleep apnea screening in the large population. A deep learning model was developed to classify the normal and apnea events in a single time-series signal exported from the US National Institute of Health (NIH) sponsored database. Our challenge was to train the model with imbalanced dataset between normal and abnormal respiratory events. Three different methods, Synthetic Minority Over-sampling Technique (SMOTE), Random Under-Sampling (RUS), and the Class Weights (CW) were chosen to improve the model performance over the original data on five selected signals from polysomnographic dataset. The binary classification outputs were evaluated by four metrics. Our results showed (1) Matthews Correlation Coefficient was highest (
$MCC= 0.1385$
) in the Class Weights method on the nasal airflow signal. (2) Cohen's Kappa score, was highest (
$k= 0.0819$
) in SMOTE technique on the abdominal signal, followed by the Class Weights method on the abdominal signal (
$k= 0.0687$
) and RUS technique on nasal airflow signal (
$k= 0.0441$
). (3) F1-score was highest (
$F1= 11.89\%$
) in SMOTE technique on the abdominal signal, followed by the Class Weights method on nasal airflow signal (
$F1 = 11.17\%$
) and RUS technique on nasal airflow signal (
$F1= 9.16\%$
). The findings suggest that the Class Weights method on nasal airflow and the Class Weights method on abdominal signal were the two combinations to be used in the DL model.