Automated Scoring of Neonatal Sleep by

This paper presents an automated sleep scoring system for neonates using hidden Markov models (HMMs) derived from features extracted from electroencephalogram (EEG) signals. The automated system can be used to score neonatal sleep in real-time and such an automated system makes it possible for a clinician to monitor a patient without an expert sleep scorer present to interpret the physiological signals. It can also be used to speed up the time-consuming process of manually scoring EEG data. In the system developed in this work, only features derived from the EEG are necessary thus making data collection easier than if a full polysomnogram (PSG) was required. Contact Information: Kenneth A. Loparo EECS Department Case Western Reserve University 10900 Euclid Ave. Cleveland, Ohio 44106-7071 Phone: (216) 368-4115 Fax: (216) 368-3123 e-mail: [email protected] 1.0 Introduction: Patterns in polysomnograms collected from sleeping neonates have clinical importance. It takes a trained electroencephalograher a long time to score these signals into a set of sleep state codes that can be more easily interpreted, and there is usually disparity in the results of the scoring between physicians. This paper presents an automated sleep scoring system for neonates based on features derived from EEG signals and hidden Markov models that can be used to score sleep in real-time. By eliminating the human factor, it would also serve as a more consistent approach to scoring. This system only requires EEG signals as input. Other physiological signals may contain useful information, but were found to be redundant in the scope of this study. The data used in this study was a subset from a sleep study performed in Pittsburgh. When scored by a physician, epochs from a neonatal sleep PSG are assigned labels based on the neonatal sleep scoring. The sleep cycle starts with Mixed frequency Active Sleep (MAS) and usually follows the sequence MAS-HVS-TA-LVI (MHTL) and then repeats. Where, High Voltage Slow (HVS), Trace’ Alternant (TA) and Low Voltage Irregular (LVI) are the other sleep states. Some of the epochs may also be scored as indeterminate when there is no clear indication of a particular sleep state or scored as transitional when the PSG epoch contains an intermediate pattern between two sleep states. The system described in this paper works by extracting features from the EEG signals to create an observation sequence. Each EEG channel is associated with activity from a section of the brain identified by the 10-20 electrode placement system. The observation sequence is classified by a hidden Markov model where each state in the HMM corresponds to an individual sleep state, e.g. MAS, or to either active sleep (MAS and LVI) or quiet sleep (HVS and TA). The remainder of this paper is organized as follows: Section 1.1 discusses previous work; Section 2 discusses the different methods used in developing this system; Section 2.1 discusses feature extraction methods; Section 2.2 discusses HMMs; Section 2.3 discusses system structuring; Section 2.4 discusses the final implementation and parameter selection; Section 3.0 presents the results; Section 4 contains the conclusions and future work. 1.1 Previous Studies Automated sleep staging systems have been implemented for adults and neonates, and they work to varying degrees, but they are far from replacing experienced electroencephalographers. These systems can effectively be used to provide a first pass to scoring which can save a scorer considerable time. Even with an existing standard for sleep stage scoring, this task is still largely subjective, and there are many ambiguous instances that different electroencephalographers disagree upon. The scoring rules only provide a guideline for scoring and for this reason, one of the main problems with automated sleep staging systems is that they show the best results in the sleep lab in which they were developed One method is to use a rule based approach. Principe et al. [12] used the Dempster-Shafer (DS) theory of evidence and feature extraction techniques to implement a set of scoring rules for an automated sleep scoring system in adults. The D-S implementation uses another type of classifier and takes the place of the HMM used in this study. They achieved 84% agreement with a physician’s scores and many of the disagreements could be explained by state transitions and confusion between two specific sleep states. Accardo et al. [1] investigated fractal dimension as a feature to be used in scoring sleep states in neonates. They found that the value of the fractal dimension, D, was cyclic with the sleep cycle and that there were differences in the value of D with respect to the topology of the scalp. These topological differences are very similar to the topological differences found in this study regarding the values of the AR coefficients derived fro the EEG time series. No classification results were presented and the computation of the fractal dimension requires many minutes of data which may prevent it from being a viable feature in a real-time system. Galhanone et al. [6] found that MAS and LVI are highly intermingled in the discriminant space. Their study used only four recordings but successfully constructed a classifier to detect MAS, LVI, and HVS with 66% agreement. Flexer et al. [5] used a hidden Markov model, two channels of EEG and a channel of EMG to create a sleep scoring system for adults. They obtained over 80% agreement for wakefulness and deep sleep, but less than 30% for the other adult sleep states. Scher et al. [15] developed a sleep staging system for neonates that also used the Pittsburgh data set, and that could successfully classify, awake, quiet sleep, and active sleep, and showed differences between the full-term and preterm groups by gestational age. This system was based on a discriminant analysis of individual recordings, whereas the approach used in this study is generalized so that the system will perform well on most arbitrary recordings. On of the methods used in this study to combine features into scores involves the nonlinear technique of system structuring [18], which is similar to discriminant analysis in that it can be used to combine features to generate a reduced feature set. Most of the sleep studies reported in the literature didn’t use as large of a data set as was used in this study and also only used a couple of channels of EEG. In this study, all of the EEG channels available were used, but only a subset was shown to be necessary. Other studies have been done on classifying states that are not related to sleep using EEG, such as recognizing when certain tasks are being performed. These studies have used various sources of data from the raw signal to extracted features, and classifiers ranging from variations of HMMs to artificial neural networks. For more information and examples refer to [7], [8], [10], [16], and [22]. In this study, the neonates were divided into two cohorts based on their gestational age and two different HMMs were used for classifying the observation sequences. If the gestational age was less than thirty-eight weeks, then the preterm HMM was used. Otherwise, the full-term HMM was used. Each HMM has the same topology and input features, but the model parameters are different. This system was tested on a total of forty-six studies between the two groups. With the approach taken, it has been shown that including other physiological signals in addition to EEG does not generally improve the classification results. Expert scorers generally agree about eighty percent of the time and in this regard, the correct classification rate of our automated system is comparable. 2.0 Methods The automated sleep scoring system for neonates described in this paper uses features extracted from spontaneous EEG signals. Determining which features are useful for classification is a difficult problem. The features were selected by a combination of visual inspection and using a new technique developed in our research group, system structuring. HMMs were used to develop the classifier because of their ability to probabilistically model the temporal characteristics of the sleep cycle. 2.1 Feature Extraction Many different feature extraction techniques were used in this work and these features were used to varying degrees for classifying a given epoch of EEG data. Since our interest is in the development of a real-time system, it must also be possible to implement the feature extraction processes in real-time. When multiple feature types are acquired at different rates, interpolation is used to create an observation sequence where each feature is acquired at the same rate. The different types of feature extraction techniques used in this system are described next. 2.1.1: Frequency Features One of the most common analysis tools is the Fourier transform. For a discrete-time signal, because the DFT assumes that the signal is of finite duration or periodic, a window is used to extract an epoch from a time series of data to be processed. To reduce the impact of edge effects from the windowing process, a Hamming window was applied to the data. The DFT data is aggregated into one Hz frequency bins and the energy in each one Hz bin is used as a separate feature. The window of data must be of reasonable length so that the underlying pattern in the signal becomes apparent with the desired frequency resolution, but not so large that there will be loss of information because of the nonstationarity of the signal. The frequency features tend to be robust to the size of the time window in this application as long as the time window stays constant throughout and the length of the window is chosen appropriately. The window size chosen for this study was twenty seconds. There are many features that can be derived from the DFT; the magnitude, phase, real part, imaginary part, power spectrum, log power spectrum, normalized power spectrum, spectral edge, mean frequency, etc. Figure 2-1 shows the magnitude spectrum versus time and the histogram of the 2-3Hz frequency bin for channel Fp1-C3 from the training set. The x% spectral edge is the frequency at which x% of the spectrum is contained at lower frequencies. Features used for this system include the 10% spectral edge through the 90% spectral edge at 10% increments for both the magnitude and the power spectrum. The 50% spectral edge is also known as the median frequency. Analogous to spectral edge is a feature that we refer to as the gravity edge. The x% gravity edge is defined to be the frequency at which x% of the area moment from the spectrum is at lower frequencies. The 50% gravity edge is also known as the mean frequency. This system uses the 10% gravity edge through the 90% gravity edge at 10% increments for both the magnitude and the power spectrum as features. The mean frequency from EEG signals has been successfully used as a feature in HMMs for detecting drowsiness [7]. The 90% spectral edge has also been shown to be cyclic in adults [20]. The feature versus time plot of Figure 2-2 for the mean frequency shows reasonably good separation between the classes. 2.1.2: Hjorth Parameters Another set of features derived from the EEG signals that are useful for distinguishing between the sleep states of neonates is the Hjorth parameters. The Hjorth parameters are shown in Equations 2-3. The activity is the signal variance and if the signal is chaotic, then the variance may not provide sufficient information about the signal for classification. The mobility measures the spread of the changes in the signal compared to the spread of the signal. The complexity is a measure of how complicated the signal is. There are other ways to measure signal complexity using measures such as the Lyapunov exponent or the correlation dimension. These methods are computationally expensive, require large amounts of data, and make assumptions that the data is stationary over large intervals. Because the mobility and the complexity measures involve the derivative operator, a low-pass filter should first be applied to the data signal.