ADAPTIVE NEURO-FUZZY INFERENCE SYSTEM FOR AUTOMATIC SLEEP MULTISTAGE LEVEL SCORING EMPLOYING EEG, EOG, AND EMG EXTRACTED FEATURES

Abstract

A new system for sleep multistage level scoring by employing extracted features from twenty five polysomnographic recording is presented. For the new system, an adaptive neuro-fuzzy inference system (ANFIS) is developed for each sleep stage. Initially, three types of electrophysiological signals including electroencephalogram (EEG), electrooculogram (EOG), and electromyogram (EMG) were collected from twenty five healthy subjects. The input pattern used for training the ANFIS subsystem is a set of extracted features based on the entropy measure which characterize the recorded signals. Finally an output selection subsystem is utilized to provide the appropriate sleep stage according to the ANFIS stage subsystems outputs. The developed system was able to provide an acceptable estimation for six sleep stages with an average accuracy of about 76.43% which confirmed its ability for multistage sleep level scoring based on the extracted features from the EEG, EOG and EMG signals compared to other approaches.

INTRODUCTION

Rechtschaffen and Kale (R&K) (1968) recommendation (or standard) classifies the human sleep into two main stages: rapid eye movement (REM) and non-rapid eye movement (NREM). They further divide the NREM stage into various stages including stage one, two, three, and four (S1, S2, S3, and S4), respectively. R&K presented different rules to determine the stage of human sleep. These rules are based on the characteristics of electrophysiological signals including electroencephalogram (EEG), electrooculogram (EOG), and electromyogram (EMG). The main characteristics of the EEG signal are alpha, beta, theta, delta, sleep spindles, and K-complex. These characteristics are categorized based on different frequency bands and amplitudes. If the alpha wave presents more than half of the studied window, it indicates that the human is not sleeping and the state is classified as wakefulness state (W). Stage one is characterized by low voltage, mixed frequency EEG without rapid eye movement. Stage two is defined by the presence of the K-complex signal along with sleep spindle. If delta waves are present in less than half of the epoch, stage three is reported; where as if delta waves are present in more than half of the studied window, then stage four is defined.

Sleep stage scoring is the process of applying R&K rules on a recorded electrophysiological signal in order to identify the stage of sleep the human is currently passing. This classification is important in studying sleep disorders. The manual process of sleep stage scoring is done based on 30-second epochs, where sleep specialists use R&K rules to determine the sleep stage of each epoch. This manual procedure is time consuming and subjective, thus leading to high inter-agreement errors between specialists during the scoring procedure (Japan Sleep Society 1999).

The process of automatic sleep-stage scoring involves two main steps; the first step is extracting different features from the electrophysiological signals that characterize the human sleep. These features usually indicate the presence of different signals such as K-complex, sleep spindle, alpha, and beta waves. Before extracting the features, different smoothing and filtering techniques are applied (Hanaoka, Kobayashi, and Yamazaki 2002). Many features were extracted using frequency domain analysis such as the spectral power of frequency bands (Fell et al. 1996; Ferri et al. 2001). Other methods involve the usage of time domain analysis (Japan Sleep Society 1999). Recent works applied Wavelet approach using the time-frequency domain analysis (Oropesa, Cycon, and Jobert 1999; Kiymika, Akin, and Subas 2004; Liu, Zhang, and Yang 2002; Fraiwan, Khaswaneh, and Lweesy 2009). The second step is the process of applying different classification techniques to identify the sleep stage. Wide range of classification techniques were used, such as linear discriminate analysis (Molinari, Dumermuth, and Lange 1984), neural networks (Schaltenbrand, Lengelle, and Macher 1993; Schaltenbrand et al. 1996; Roberts and Tarassenko 1992; Shimada, Shiina, and Saito 1998; Lalitha and Eswaran 2007; Tagluk, Sezgin, and Akin 2009; Kurt et al. 2009; Vuckovic et al. 2002; Acir and Guzelis 2004; Fraiwan, Khaswaneh, and Lweesy 2009), Neuro-Fuzzy logic (Pohl and Fahr 1995; Yildiz et al. 2009), tree classification (Kubat, Pfurtscheller, and Flotzinger 1994; Fraiwan, Khaswaneh, and Lweesy 2009), Hidden Markov Models (Flexer et al. 2000; Flexer and Dorffner 2002), and case-based reasoning approach (Park et al. 2000). Table 1 provides an illustration of the available state of the art systems reported in literature for sleep stage detection.

TABLE 1 Available Systems Reported in Literature for Sleep Stage Detection

Authors	Classified stages	Classifier approach	Results
Watanabe and Watanabe (2004)	– Non-REM stages– REM stage– Wake	Mathematical modeling	82.6%38.3%70.5%
Šušmáková and Krakovská (2008)	Wake and four sleep stages	Fractal exponent and fractal dimension	Classification error 23%
Zoubek et al. (2007)	Sleep/Wake	Supervised classification methods (Bayes rule-based classifiers and Multi-layer Perceptron)	About 80%
Pinero et al. (2004)	Wake, REM and four sleep stages	Fuzzy model	78–84%
Flexer, Gruber, and Dorffner (2005)	Wakefulness, deep, and REM sleep	Hidden Markov model	Around 80%
Louis, Lee, and Stephenson (2004)	– Wake and Sleep– Wake, REM and NREM	Computer based steps algorithm	90.1%87.5%
Fraiwan, Khaswaneh, and Lweesy (2009)	Wake, REM and four sleep stages	Wavelet packets decomposition	About 74%
Tagluk, Sezgin, and Akin (2009)	REM and four sleep stages	Artificial Neural Networks	About 74.7%
Saastamoinen et al. (2006)	– Wake– Light sleep– Deep sleep– REM	Computer based spectrum approach with two parameters^+: window length & band frequency.	69–76.6%^+41–43.1%^+78–93.1%^+44.8–51.4%^+

In this article, an automated sleep multistage level classification model is developed using an adaptive neuro-fuzzy inference subsystem (ANFIS) for each sleep stage. Each subsystem combines the fuzzy logic system expert knowledge with the neural networks learning abilities to exploit its functionality based on the ANFIS structure. Each ANFIS subsystem is trained with features extracted from EEG, EMG, and EOG signals to determine its corresponding stage; W, 1, 2, 3, 4, or R, corresponding to wake, sleep stages: one, two, three, four and REM, respectively. The final estimated sleep stage from all subsystems is determined through an output selector subsystem. The latter combines all results from all subsystems and provides the best classification for the applied features. To achieve higher sensitivity and accuracy in sleep multistage classification, the system employs a parallel architecture, where each level corresponds to one sleep stage subsystem.

SYSTEM ARCHITECTURE

To sustain a high sensitivity and accuracy for all sleep stages, parallel system architecture is proposed as shown in Figure 1.

FIGURE 1 The overall system architecture.

Signal Processing Unit

Training data by feeding the collected patients' data (EEG, EOG, and EMG) to a signal processing unit. The extracted features from the signal-processing unit are used to train the ANFIS subsystem models for each sleep stage. The generated model is used to classify further sleep stages after training. The signals features are extracted from five electrophysiological channels. These five channels are allocated as follows, two EEG channels, the left and right eye EOG channels, and the chin EMG channel. Each of the signals recordings are divided into time segments of 30 seconds. Figure 2 shows different recorded epochs corresponding to sleep stages: W, 1, 2, 3, 4, and R, respectively.

FIGURE 2 Different recorded EEG, EOG, and EMG epochs corresponding to sleep stages: a – W, b – 1, c – 2, d – 3, e – 4, and f – R.

The signals from the two EEG channels are recorded using C3A2 and C4A1 electrodes. Features are extracted from both EEG channels based on the frequency range as shown in Table 2. Each feature represents the corresponding activity in the EEG signal: alpha wave, beta wave, theta wave, delta wave, K-complexes, and sleep spindles. The features calculation is carried out using the power spectrum estimation method of the EEG's 30-second time segment. The power spectrum estimation method is based on computing the power spectral density (PSD) using the Periodogram (Stoica and Moses 1997). The power spectral density S _x (f) of a random time signal x(t) can be calculated as (Stoica and Moses 1997; Jackson 1989):

The power (P) in x(t) in a range of [f ₁ f ₂] can be evaluated as follows:

For each EEG event, different f ₁ and f ₂ values are substituted in Equation (2). The frequency values in Table 2 were utilized in the computations. The relative power of each event is used to detect its presence in the EEG signal. In addition to the power spectrum estimation, the amplitude variation of the EEG signal is calculated. Moreover, the level of the amplitude variation is calculated using the standard deviation of each epoch. Finally, the third order linear prediction coefficients of the EEG are estimated (Quinlan 1993).

TABLE 2 Frequency Range for EEG Events and Waves

Event	Frequency range (Hz)
K-complex	0.4–1.5
Delta	0.5–4
Theta	4–8
Alpha	8–13
Sleep spindle	12–14
Beta	13–25

The correlation coefficient between the left and the right EOG are calculated to detect the state of the EOG. Furthermore, the total relative power for the ranges of [0.15–0.45] Hz, and [0.5–1.2] Hz are computed as well. The tone of the chin (EMG) is extracted as another feature, which can be exploited to differentiate between the REM and the wake stage. The EMG tone is evaluated by dividing the EMG signal into sub-epochs, each one has half second duration. The average of variances of the sub-epochs is then estimated.

A features selection process comes normally as the last step before the classification procedure. Some features may be redundant and tend to degrade the accuracy of the classification procedure; therefore, a features selection procedure is implemented. The gain ratio entropy approach is used for input features selection. The gain ratio is evaluated by measuring the gain ratio with respect to the class as (Witten and Frank 2000):

where H specifies the entropy as:

where p denotes the probability mass function of X. Table 3 shows the ranked attributes based on the entropy measure as in Equation (4). The input features are selected based on attributes with higher entropy values.

TABLE 3 Ranked Attributes Based on Entropy

Entropy	Attributes	Entropy	Attributes	Entropy	Attributes
0.1415	LPC2A	0.1140	TotalPowerB	0.0815	EMGStd
0.1415	LPC2B	0.1017	ThetaA	0.0810	AlphaB
0.1398	DeltaA	0.0968	EOGRP2	0.0783	LPC4B
0.1367	BetaA	0.0912	AlphaA	0.0783	LPC4A
0.1322	DeltaB	0.0906	SpindleA	0.0649	EOGLP1
0.1317	LPC3B	0.0899	KComplexB	0.0520	EMGVar
0.1317	LPC3A	0.0896	KComplexA	0.0519	EOGRP1
0.1259	BetaB	0.0868	ThetaB	0	LPC1B
0.1238	TotalPowerA	0.0868	EOGLP2	0	LPC1A
0.1199	EOGCorr	0.0818	SpindleB	–	–

ANFIS Subsystem Stages Model

The proposed new system is based on a parallel architecture. Each sleep stage is modeled by an Adaptive Nero-Fuzzy Inference System (ANFIS), which corresponds to one level of the parallel architecture. Such system combines the fuzzy logic system expert knowledge with the neural networks learning abilities to exploit its functionality (Jang 1993). The ANFIS architecture of each sleep stage consists of five-layered feedforward neural network, as shown in Figure 3. These five layers implement a Sugeno fuzzy expert system (TSK) of the following form (Jang 1993):

where the input membership functions linguistics variables are A and B and the output parameters a _k are the design parameters to be tuned using the neural networks learning abilities for each rule.

FIGURE 3 The ANFIS architecture of each sleep stage.

The input layer provides the subsystem with the extracted features. Layer 1 is an adaptive fuzzification layer. The output of each node Equation (6) represents the matching degree with its membership function (μ _Ai ), that corresponds to a linguistic label A _i (e.g.: small, medium, or large) for the given input (x _i ) (Jang 1993).

A Gaussian bell-shaped membership function (μ _Ai ) (Equation (7)) is selected to map the input value (x _i ) into its equivalent fuzzy value within the interval [0 1]. Four (n) linguistic labels are used to describe the five inputs (m), which are the extracted features {m _i , α _i , β _i } are the set of premise modifiable parameters, which are tuned during the learning phase (Jang 1993). For instance, the initial membership functions for the five inputs of the stage-W subsystem are illustrated in Figure 4. Also, Figure 4 shows the universe of discourse for each input.

FIGURE 4 Initial membership functions of the five inputs for the stage-W subsystem.

The firing strength of each rule (r) is determined in layer 2. Each node's output is the normalized product (T-norm operator) of its input signals as in Equation (8):

Layer 3 is an adaptive defuzzification layer, where each node's output is a normalized weighted output of the first order fuzzy TSK IF-THEN rule of the following form (Jang 1993):

The set of the consequent parameters (a _i ) for each rule are tuned during the training phase. Layer 4 completes the defuzzification process by summing all the normalized weighted outputs from the previous layer, which define a weighted average single defuzzified output for subsystem k (Jang 1993):

Training ANFIS Subsystems

During the training phase, each ANFIS subsystem tunes the entire premise and the consequent modifiable parameters for the adaptive nodes, such that the ANFIS minimizes the error between the structure output and a predefined training dataset. This set contains inputs/output pairs of the desired network behavior provided to the system by an expert. These parameters are initially generated by grid partitioning for the input space based on the numbers of the inputs of each subsystem and the membership functions (Jang 1993). For this purpose, a hybrid training algorithm is used, which combines the least square method with the gradient descent method (Jang 1993). The least square method is used to tune the consequent modifiable parameters during the forward pass while the gradient descent method is used to tune the premise modifiable parameters during the backward pass (Jang 1993). The ANFIS features for each stage subsystem are shown in Table 4. For instance, the membership functions for the five inputs of the stage-W subsystem after training are shown in Fig. 5 with the universe of discourse for each input.

FIGURE 5 Membership functions for the five inputs of stage W subsystem after training.

TABLE 4 Features for the ANFIS Subsystem Structure for Each Stage

Number of nodes:	2096
Number of linear parameters:	1024
Number of nonlinear parameters:	60
Total number of parameters:	1084
Number of training data pairs:	9266
Number of fuzzy rules:	1024
Number of layers:	5
Input layer nodes:	5
Output layer nodes:	1
Training approach:	Hybrid algorithm(combines the least square method with the gradient descent method)

Output Selection Subsystem

The automatic sleep stage level scoring is achieved according to the ANFIS stage subsystems output, which satisfies the following criterion:

where ζ is an output vector resulted from applying the extracted features x to all ANFIS stages subsystems.

EXPERIMENTAL RESULTS AND DISCUSSION

To evaluate the performance of the new proposed system, experiments are performed on 25 digital polysomnographic recordings. These recordings were collected from adult subjects with suspected sleep-disordered breathing. Subjects were randomly selected over a 6-month period (September 2002 to February 2003) from patients referred to the Sleep Disorders Clinic at St. Vincent's University Hospital, Dublin, for possible diagnosis of obstructive sleep apnea, central sleep apnea, or primary snoring (Goldberger et al. 2000). Subjects had to be above 18 years of age, with no known cardiac disease, no autonomic dysfunction, and they could not have been on medication known to interfere with heart rate (Goldberger et al. 2000).

Polysomnograms signals were collected using the Jaeger-Toennies system (Goldberger et al. 2000). Signals used are: EEG (C3-A2), left EOG, right EOG, and submental EMG.

For the all 25 subjects, sleep stages are scored by the same experienced sleep technologist according to standard R&K rules. Fifteen recordings are used for training the model and the rest ten recordings are used for system evaluation after training (Goldberger et al. 2000). The set of the first fifteen recordings was used for training the sleep stage ANFIS subsystem. The same training set was also used to obtain the optimal number of membership functions to represent the universe of discourse for each extracted feature to achieve an acceptable generalization for the developed system. For this purpose, a thorough study is conducted on the training data based on the performance of the proposed system upon the following statistical indices (Kiymik, Akin, and Subasi 2004; Yildiz et al. 2009; Lalitha and Eswaran 2007):

and

where TP, TN, FP, and FN are the level classification being: true positive, true negative, false positive, and false negative, respectively. The response of the proposed system for the fifteen recordings based on the above indices and different number of input membership functions is demonstrated in Tables 5-7, and in Figure 6. As it can be concluded from the tables, the optimal number of membership functions for the sleep stage ANFIS subsystem is selected to be four, where the overall accuracy is improved from an average of approximately 87.27%, 88.53% and 89.03% for two, three and four membership functions, (Figure 6) respectively. As shown in Figure 7, the same trend is achieved through the other indices. For instance, the sensitivity is improved to an average of near 93.79% by using four membership functions. Selectivity and the accuracy are improved to an average of 94.55% and 93.70%, respectively.

FIGURE 6 Overall accuracy using two, three, and four membership functions (MF) to represent the input features.

FIGURE 7 Sensitivity, selectivity, specificity and accuracy for sleep stages subsystems using two, three and four membership functions (MF) to represent the input features.

TABLE 5 The Performance of the System Using Two Membership Functions

Number of MF = 2	Stage W	Stage 1	Stage 2	Stage 3	Stage 4	Stage R
Sensitivity (%)	95.88	89.74	87.79	92.74	96.94	95.08
Selectivity (%)	92.93	95.15	86.65	100.0	90.03	95.29
Specificity (%)	91.65	94.84	83.93	100.0	87.66	94.92
Accuracy (%)	93.77	92.29	85.86	96.37	92.30	95.00
Overall accuracy (%)	89.37	85.82	77.34	92.74	87.55	90.82

TABLE 6 The Performance of the System Using Three Membership Functions

Number of MF = 3	Stage W	Stage 1	Stage 2	Stage 3	Stage 4	Stage R
Sensitivity (%)	95.96	90.68	89.86	92.86	96.22	95.96
Selectivity (%)	93.62	94.75	88.20	99.84	91.63	97.15
Specificity (%)	92.46	94.36	85.53	99.85	89.71	96.90
Accuracy (%)	94.21	92.52	87.70	96.36	92.97	96.43
Overall accuracy (%)	90.07	86.35	80.22	92.74	88.44	93.33

TABLE 7 The Performance of the System Using Four Membership Functions

Number of MF = 4	Stage W	Stage 1	Stage 2	Stage 3	Stage 4	Stage R
Sensitivity (%)	95.77	91.35	90.06	93.13	96.15	96.26
Selectivity (%)	94.76	94.54	89.05	99.72	92.04	97.18
Specificity (%)	93.83	94.14	86.55	99.72	90.23	96.92
Accuracy (%)	94.80	92.75	88.31	96.43	93.19	96.59
Overall accuracy (%)	90.96	86.81	81.08	92.89	88.78	93.64

To sustain the subsystems generalization for each sleep stage, the number of membership functions is not increased further. In such case, the ratio between the set of training data to the set of modifiable parameters for each ANFIS subsystem decreases and affects the generalization of the developed system. This is due to notable increase in ANFIS modifiable parameters.

After training, a set of 10 recordings is utilized for system evaluation, the same statistical indices: sensitivity, selectivity, specificity, and accuracy are used during this phase. The results obtained from this phase are tabulated in Table 8. As it can be analyzed from the table and after comparing with experts knowledge, the developed system is able to provide an acceptable estimation for all sleep stages with average percentage accuracy close to 76.43%, and an average percentage sensitivity of just about 90.02%.

TABLE 8 The Performance of the System

	Stage W	Stage 1	Stage 2	Stage 3	Stage 4	Stage R
Sensitivity (%)	92.09	83.73	81.95	94.66	97.19	89.79
Selectivity (%)	61.92	66.38	67.54	60.81	59.81	63.00
Specificity (%)	53.76	64.98	56.93	60.72	55.76	59.61
Accuracy (%)	85.65	74.36	69.44	77.69	76.48	74.70

The results of the new developed system are compared to other available systems reported in literature for sleep stage detection, as shown in Table 1. The system that has been developed in (Tagluk, Sezgin, and Akin 2009) is based on a multilayer artificial neural network (ANN) for sleep stage estimation, stages: 1, 2, 3, 4, and REM using EEG, EMG, and EOG signals. The system was able to provide an average percentage accuracy of 74.7% after training. Comparing the new developed system output with R&K rules (Rechtschaffen and Kales 1968; Tagluk, Sezgin, and Akin 2009; Watanabe and Watanabe 2004), it shows an acceptable performance and agreement for 70.5% wakefulness, 38.3% REM and 82.6% NREM stages an average percentage of 63.8%. When comparing the new developed system response with classification method for automated sleep stage scoring based on a single EEG recording using wavelet packet decomposition (Watanabe and Watanabe 2004), this approach was able to achieve an average accuracy about 74% during testing phase. Therefore, the new developed system based on parallel ANFIS subsystems is able to improve the accuracy by combining the fuzzy logic system expert knowledge with the neural networks learning abilities to exploit its functionality.

CONCLUSION

This article presents an efficient, automated, online approach for sleep multistage level scoring employing extracted features form polysomnographic recording. For the new system, an adaptive neuro-fuzzy inference subsystem (ANFIS) is developed for each sleep stage by combining the fuzzy logic system expert knowledge with the neural networks learning abilities to utilize their functionalities.

The extracted features, which characterize the electrophysiological signals including electroencephalogram (EEG), electrooculogram (EOG), and electromyogram (EMG), were fed into each stage of the ANFIS subsystem as the input pattern for training and evaluating the new approach. The interesting performance for the new automated approach compared with other proposed approaches, confirmed its ability for online multistage sleep level scoring based on the extracted features form the EEG, EOG, and EMG signals.