Effect of Resampling Techniques on Deep Learning Model Training in Sleep Apnea Classification

Abstract

This study is using deep learning model to classify the respiratory events of Sleep Disordered Breathing (SDB) data. Our pilot results showed the missed identification in some classes even the total accuracy is high. This is the result of unbalanced training dataset given to the model. Two different resampling techniques; Synthetic Minority Over-sampling Technique (SMOTE) and Random Under-Sampling (RUS), were introduced to balance the data. One hundred overnight nasal airflow signals were randomly selected from NIH funded polysomnography database. They were used to train and test these two algorithms with Bi-directional Long Short-Term Memory (Bi-LSTM) model. The results showed greater agreement index when compared between with and without data resampling process. However, SMOTE in sum performed better than RUS (93.72% vs 70.01% in overall accuracy and 0.91 vs 0.55 in Cohen’s kappa). It demonstrates that the over-sampling technique is more powerful than under-sampling one. Other resampling techniques will be investigated to make the robust conclusion.