Artificial neural network based arousal detection from sleep electroencephalogram data

Electroencephalographic arousals are characterized by a sudden shift in electroencephalogram (EEG) frequency during sleep. Occurrence of arousals causes improper sleep which is the main reason of day-time sleepiness. The arousal during the sleep is detected by analyzing a multimodal multichannel Polysomnographic (PSG) recordings. It is a time consuming task to analyze the recording manually and requires a lot of patience. Hence automation of the process is required. This task becomes difficult because of the presence of a lot of events in time in relation to various bio-medical signals available through Polysomnographic (PSG) recordings. In this paper we present a method to detect the arousals in sleep automatically. Firstly two sets of electroencephalogram (EEG) channels C
₄
/M
₁
and C
₃
=M
₂
and an electromyogram (EMG) channel are chosen for preprocessing. Then the electroencephalogram (EEG) signals are passed through a bandpass filter of 8-30 Hz in order to extract the signal containing only the alpha and beta frequency components. The events are detected when the Power spectral density of this filtered signal crosses a threshold of zero. When an event is detected the relevant features from Alpha, Beta, Theta, Delta, and sigma waves are extracted. Similarly when electromyogram (EMG) signal crosses a threshold of zero events are detected and relevant features are extracted for the corresponding events. Now all these features are grouped together along with the corresponding labels and used as inputs to the Artificial Neural Network (ANN) Classifier to detect the presence of arousals. The novelty of this work relies on using both Hjorth and Power Spectral density difference spectrum features leading to improved accuracy than either of them alone. Considering 10 overnight sleep recordings We recorded an average sensitivity of 0.93262, average specificity of 0.91387, average precision of 0.91693 and average Area under Region of Convergence curve (AUC) of 0.92328 (showing a good measure of overall performance).