Automatic recognition of alertness level by using wavelet transform and artificial neural network

Abstract

We propose a novel method for automatic recognition of alertness level from full spectrum electroencephalogram (EEG) recordings. This procedure uses power spectral density (PSD) of discrete wavelet transform (DWT) of full spectrum EEG as an input to an artificial neural network (ANN) with three discrete outputs: alert, drowsy and sleep. The error back propagation neural network is selected as a classifier to discriminate the alertness level of a subject. EEG signals were obtained from 30 healthy subjects. The group consisted of 14 females and 16 males with ages ranging from 18 to 65 years and a mean age of 33.5 years, and a body mass index (BMI) of 32.4 ± 7.3 kg/m². Alertness level and classification properties of ANN were tested using the data recorded in 12 healthy subjects, whereby the EEG recordings were not used been used to train the ANN. The statistics were used as a measure of potential applicability of the ANN. The accuracy of the ANN was 96 ± 3% alert, 95 ± 4% drowsy and 94 ± 5% sleep. The results suggest that the automatic recognition algorithm is applicable for distinguishing between alert, drowsy and sleep state in recordings that have not been used for the training.

Introduction

The aim of this study was to establish a method for processing input data from a full spectrum of (EEG) recordings by the use of an artificial neural network (ANN) that distinguishes between alert and drowsy states in arbitrary subjects by the use of DWT processed EEG signals.

EEG distinguishes between states of vigilance, that is, wakefulness and sleep, and to some extent between the ‘levels’ of vigilance within a state. The EEG frequency spectrum is subdivided into δ (1–4 Hz), θ (4–8 Hz), α (8–13 Hz), β (13–30 Hz) and γ (>30 Hz) frequency ranges. Within NREM sleep, δ power (slow wave power) indicates the intensity of sleep and represents the need for sleep. During wakefulness, α and θ frequencies in the awake state EEG are of particular interest for research on sleepiness. During active wakefulness (with eyes open), α power is usually low unless the subject is severely fatigued. However, in resting conditions (with eyes closed), α power is also high when the subject is fully rested. During the transition from resting conditions, with eyes closed, to sleeping a gradual reduction of α power and a gradual increase in θ power occurs. Reduced α power and increased θ power during resting awake periods, with eyes closed, may thus indicate a high motivation for sleeping. Indeed, it was found that subjective sleepiness during awake periods correlates negatively with α power and positively with θ power in the awake EEG during prolonged wakefulness.

Spontaneous electrical brain activities, that is EEG signals, are dynamic, stochastic, non-linear and non-stationary (Guler et al., 2001, Herrmann et al., 2001, Vuckovic et al., 2002, Peters et al., 1998). The EEG recordings depend on the location of the electrodes, their impedance and the state of alertness. In addition, the EEG recordings vary substantially between healthy subjects. Extensive expertise is required to visually interpret the EEG recordings in order to isolate and identify characteristic information from a large amount of data. A computerized analysis of the EEG recordings aims to facilitate the time-consuming and difficult visual inspection and automatically extract characteristic features of brain activity. A computer-assisted EEG classification of drowsiness has been analyzed in several studies (Anderson et al., 1995, Doghramji et al., 1997, Gevins and Smith, 1999, Jung et al., 1997, Khahill and Duchene, 1999, Principe et al., 1989, Tsoi et al., 1994, Wilson and Bracewell, 2000). The classification was based on a spectral analysis of EEG recordings (Doghramji et al., 1997, Jung et al., 1997) and showed that a limited number of electrodes and spectral analysis of characteristic bands could be used as a classifier. More recently, some studies (Jung et al., 1997, Peters et al., 2001) concentrated on detecting the information on drowsiness available from a full EEG spectrum. Principe et al. (1989) designed a finite automaton that was capable of categorizing the sleep into seven different stages. McKeown et al. (1997) used statistical methods for the analysis of EEG signals and detection of vigilance changes. Pradhan et al. (1996) presented preliminary results for the classification of seizure activities by applying an ANN based on learning vector quantization. Kalayci and Ozdamar (1995) showed that an ANN performs better if the input and output data can be processed to capture the characteristic features of the signal (Anderson et al., 1995, Dorffner et al., 1993, Gevins and Smith, 1999, Haselsteiner and Pfurtscheller, 2000, Peters et al., 2001, Principe et al., 1989, Tsoi et al., 1994, Wilson and Bracewell, 2000). The combination of Fourier transform analysis of EEG with ANN in classifying alertness and drowsiness was previously shown to be a suitable algorithm for classifying events from raw EEG signals (Jung et al., 1997), except for specific conscious tasks. De Carli et al. (1999) worked on developing an automatic procedure for arousal detection during sleep. They tested this on a group of subjects, in different pathological conditions by using wavelet transform. The aim of this study was to develop a simple algorithm to discriminate the vigilance states, that is, wakefulness and sleep which could also be applied to real-time.

The following reasons were the basis for improving the methods of automatic detection of changes from alert to drowsy, and vice versa states: (1) clinical pre-processing of long-term recordings of wakefulness in order to select sequences of alert and drowsy states for further human inspection (Shimada et al., 2000); (2) online experiments, where timing of a stimuli for cognitive evoked potential are needed; (3) software for interactive learning (Akay et al., 1998); and (4) warning systems for detecting the drowsiness in operator rooms. The specific design requirement was applicability of the algorithm to short sequences of EEG recordings, hence plausible use in real-time. The second requirement was to develop a simple algorithm that would work on recordings that were not been used for the training of the same or arbitrary subject.