Bidirectional LSTM with MFCC feature extraction for sleep arousal detection in multi-channel signal data

Abstract

The polysomnography (PSG) can be used as a basis for judging various disorders that occur during sleep such as arousal. Arousal which means wakefulness is the common phenomena disturbing deep sleep. Since arousal appears in various forms, there are areas where research has been less advanced such as Respiratory effort-related arousal (RERA). We develop bidirectional Long Short-Term Memory (LSTM) which used Mel-frequency cepstral coefficient (MFCC) for feature extraction and trained using 13 multi-channel signals from Physionet Challenge 2018. The training model predicts arousal probability on every input data. Signals are processed with MFCC and we test a various combination of features such as the number of features and additional delta feature. Finally, top 3 models are used to construct an ensemble model which shows the best performance in our experiments. We obtain 0.898 AUC-ROC and 0.458 AUC-PR on the test data which is split from 994 training data. Performance of our model is competitive to other methods proposed in the Physionet Challenge 2018. Bidirectional LSTM makes a sequential prediction on arousal and MFCC can be applied uniformly on the signal data regardless of signal type. Therefore, we can process feature extraction efficiently without any manual approaches.