Automated detection of neonate EEG sleep stages

Abstract

The paper integrates and adapts a range of advanced computational, mathematical and statistical tools for the purpose of analysis of neonate sleep stages based on extensive electroencephalogram (EEG) recordings. The level of brain dysmaturity of a neonate is difficult to assess by direct physical or cognitive examination, but dysmaturity is known to be directly related to the structure of neonatal sleep as reflected in the nonstationary time series produced by EEG signals which, importantly, can be collected trough a noninvasive procedure. In the past, the assessment of sleep EEG structure has often been done manually by experienced clinicians. The goal of this paper is to develop rigorous algorithmic tools for the same purpose by providing a formal scheme to separate different sleep stages corresponding to different stationary segments of the EEG signal based on statistical analysis of the spectral and nonlinear characteristics of the sleep EEG recordings. The methods developed in this paper can, potentially, be translated to other areas of biomedical research.

Introduction

The paper integrates and adapts a range of advanced computational, mathematical and statistical tools for the purpose of analysis of neonate sleep stages based on extensive electroencephalogram (EEG) recordings. Past the neonate period, a major reorganization of sleep is known to occur around the 36th week of post-conceptional age, after which time, the regular ultradian (that is, cycling regularly more than once a day) sleep rhythm of the fullterm infant becomes better recognized. In particular, as the brain matures the percentage of active sleep decreases during the first 36 months after conception, and then typically stabilizes. The age-dependent level of brain dysmaturity of a neonate, which can also be affected by environmental factors such as premature birth, maternal drug addiction, etc., as well as the child’s genetic endowment, is difficult to assess by direct physical or cognitive examination. However, dysmaturity is known to be directly related to the structure of neonatal sleep as reflected in the nonstationary time series produced by EEG signals which, importantly, can be collected trough a noninvasive procedure. A noninvasive, reliable quantitative evaluation of the dynamic process of brain maturation is thus important for expanding our understanding of neuroplasticity, i.e., the interplay between genetic endowment and environmental influences on neonatal brain development.

In the past, the assessment of sleep EEG structure has often been done manually by experienced clinicians, but even they can differ in their interpretations. To the untrained naked eye the multi-dimensional signals appearing in examples obtained in different sleep stages in Fig. 1 are impossible to differentiate. Each of these figures, a typical EEG record, show signals measured by sensors located at different locations on the scalp, plus cardiac, breathing and facial muscle activity signals.

Below, we present a rigorous algorithmic tool to separate different sleep stages corresponding to different stationary segments of the EEG signal. It is our belief that the methods developed in this paper can, potentially, be translated to other areas of biomedical research which require analysis of piecewise stationary signals.

Let us now turn to some technical details. The EEG signal is nonstationary and therefore a single set of even infinite-dimensional, commonly used parameters, such as the power spectrum, cannot represent it adequately over long periods of time. Nonstationary phenomena are present in the EEG usually in the form of transient events, such as arousals, or as the alteration of relatively homogeneous intervals (epochs) with different statistical features (e.g., different amplitude, or variance), or sharp waves, spikes, or spike-wave discharges which are characteristic of epileptic EEG patterns [20].

However, it is the conventional wisdom (see, e.g. Ref. [3]) that within a given sleep stage the EEG signal is approximately stationary and this is our working assumption, that is, we treat the signal as a piecewise stationary stochastic process. We have verified the stationarity for various segments of our data, varying the length of these segments from 15  s, to 30 s, and to 1 min, and were satisfied that this assumption is correct. Fig. 2 shows typical sample autocorrelation functions (ACFs) (for definition, see, e.g. Ref. [37]) from different segments of the quiet (Left) and active (Right) sleep stages. Although there is some sample-to-sample variability in the ACFs, the plots shown in Fig. 2 are pretty close to the average ACF.

Neurophysiologists have traditionally identified, often by visual analysis of the EEG of fullterm neonates, four distinct electroencephalogram patterns during sleep [34], [30], utilizing the concept of spectral power distribution in five commonly used frequency bands: 0.5–4  Hz (Delta Waves), 4–8  Hz (Theta Waves), 8–12 Hz (Alpha Waves), 12–30 Hz (Beta Waves), and 31 and more Hz (Gamma Waves), see an example in Fig. 3 (which does not include Gamma Waves because of the relatively low resolution of our time series). An anti-aliasing low pass filter was used prior to acquisition of the signals to minimize aliasing effects.

• Mixed frequency active (MFA) sleep. This stage of sleep usually begins immediately after falling asleep and is characterized by stationary and continuous signals with most of the power spectrum spread between 0.5 and 10  Hz. The energy is concentrated in the Theta band with intermittent Delta waveforms (see, for example, Fig. 1 (Top Left)).

• Low voltage irregular (LVI). This active sleep stage is characterized by relatively lower amplitude signals at higher frequencies (see, for example, Fig. 1 (Top Right)).

• High voltage slow (HVS). This quiet sleep stage appears after mixed frequency active sleep, and is a brief state characterized by a shift in the frequency distribution to higher power in the lower (Delta) frequencies (see, for example, Fig. 1 (Bottom Left)).

• Tracé alternant (TA). This quiet sleep stage is characterized by nonstationary EEG patterns consisting of alternating broadband bursts of activity with intermittent epochs of relative EEG quiescence (see, for example, Fig. 1 (Bottom Right)).

Several methods for the automatic detection of sleep stages for adults have been developed, see, e.g., Refs. [10], [32], [5], [17]. A review of the fixed interval approach for the segmentation of EEG signals can be found in Barlow [2]. Recall, that in this approach the signal is split into fixed (as opposed to continuously sliding), say 30-s, time segments for each of which the relevant characteristics are computed individually. It corresponds to the way a neurologist scores the signals visually minute-by-minute.

For neonates, the structure of sleep is very different from that of the adults, and the corresponding classifications used by the neurologists are also different. Thus the study of such signals presents slightly different problems. Here, an automatic procedure for sleep stage scoring are reported in Levy [19], and methods for an automatic detection of the tracé alternant stage using the discrete wavelet transform are discussed in Turnbull et al. [35]. Gerla et al. [11] used the hidden Markov chain models algorithm, in combination with EM algorithm, to separate and identify the sleep stages. Krajča et al. [18] produced an automatic method for of EEG-sleep data separation into active and quiet sleep using adaptive segmentation and hierarchical cluster analysis. Gerla et al. [12] also suggested a hybrid evolutionary approach to newborn sleep stages classification. Most of the above work relies on training algorithms which use directly physicians’ scoring and also rely on signals which are filtered into different, previously selected frequency bands (Delta, Theta, etc., see Section 2 for more details).

Temporal patterns of sleep EEG in neonates, were shown to be useful for the assessment of functional brain maturation of infants that are at risk of developmental disabilities in Scher et al. [27], and Whitney and Thoman [36]; a discussion of their relevance for specific clinical syndromes such as sudden infant death syndrome can be found in Glotzbach et al. [13]. Different numerical measures of dysmaturity based on the statistical analysis of temporal patterns of neonatal EEG-sleep have been developed [28], [29]. The differences in sleep EEG organization between pre- and fullterm cohorts at matched post-conceptional ages were studied in, e.g., Refs. [26], [31].

Traditionally, as far as mathematical/statistical tools are concerned, EEG signals had been analyzed using spectral analysis methods, and much of the EEG literature is devoted to the study of power in the conventional Alpha, Beta, Delta, and Theta frequency bands mentioned above. However, more recently, methods of nonlinear dynamics (e.g., correlation dimension, and the largest Lyapunov exponent) have also been applied to analysis of EEG signals. According to some of the recent literature. Rombouts et al. [25], and Pritchard et al. [23] point out that neurophysiological mechanisms that generate the EEG signals are highly nonlinear. Fell et al. [9] have shown that the combination of spectral and nonlinear dynamic characteristics yielded better overall discrimination of sleep stages than spectral characteristics alone; that observation motivated our systematic approach to the search for optimally discriminating combinations of various characteristics of EEG signals.

In Sections 2 Characteristics and methods used in analysis of sleep EEG patterns, 3 Significant characteristics for neonate sleep EEG analysis, 4 Sleep stage separation via change-point detection we propose a formal algorithm for separation of a nonstationary EEG signal into stationary segments corresponding to different sleep stages. The algorithm is then applied to analyze sleep EEG patterns of pre- and fullterm neonates from the data set collected at the University of Pittsburgh. The segments of raw recordings used by us here are available at http://stat.case.edu/ayp2/EEGdat. This data set contains 21 fullterm, and 16 preterm recordings of infants of the same post-conceptional age of 40 weeks, each containing the signal from a single channel signal of EEG data sampled at 64  Hz for 2–3 h periods (see Section 3 for more details). The manual, minute-by-minute classification (what we later call “gold standard scoring”) of portions of these signals as belonging to one of the four sleep stages was provided by Dr. Mark Scher of the Pediatric Neurology Department at Case Western Reserve University.

Accepting the piecewise stationarity hypothesis discussed in Section 1.2, the first step in our approach was to devise a method to separate the EEG signals into stationary segments by estimating the transition points of the recordings; these are referred to as “change-points” in the statistical literature. Here we relied on a non-parametric statistical analysis method developed by Brodsky and Darkovsky in the 1970s, summarized in Brodsky and Darkovsky [4], and applied to macrostructural EEG-sleep characterization in Brodsky et al. [5] and Kaplan et al. [17]. Then our automatic scoring approach combined the change point detection schemes, cluster analysis, and a variety of traditional and nontraditional spectral and nonlinear characteristics in different combinations. Their effectiveness of different combinations was subsequently evaluated by calculating an agreement index which compared our algorithmic method with the “gold standard” manual scoring method. Finally, the best performing mulitvariate characteristics were then selected to implement the final classification scheme. They performed well, achieving 80–90% rate of agreement with the expert manual scores; not far from the real-life variability between analyses of independent clinicians. Indeed, similar percentages of agreement are seen between different neurologists scoring the same data, see, e.g., a study by Danker-Hopfe et al. [7], who investigated the reliability of manual sleep-EEG scores produced by eight European sleep laboratories. The choice of significant characteristics in our algorithmic scheme was validated by running the non-parametric Wilcoxon test. Also, we have found spectral and nonlinear dynamics characteristics which are significantly different for full-perm and preterm groups of neonates.

Although our approach was based on the method presented in Brodsky et al. [5], and Kaplan et al. [17], several important differences between our approach and theirs should be noted.

Firstly, their method, was developed for adult sleep signals, while ours was specifically designed for the neonate case. Secondly, their non-parametric method used filtered Delta, Theta, Alpha, Beta waves with the moving averaging over 5-s time intervals and then a diagnostic sequence was produced by finding the total power as autocorrelation function’s value at zero. In our study such high resolution was not necessary since our algorithm was designed to replace the standard manual minute-by-minute scoring system used in brain maturation studies. However, our coarser segmentation approach enabled us to use not only the standard four frequency bands but also nonlinear characteristics of the signals such as the fractional dimension, the spectral entropy, the amplitude entropy, and others. Our algorithm not only separated data into (quasi)stationary segments and clusters but also identified which one belong to the active and which one belong to the quiet sleep stages.

The studies of Refs. [11], [12], [18], also used adaptive learning algorithm for the separation of the EEG signal into homogenous segments. So, again, the advantage of our completely automated non-parametric segmentation approach, with the thresholds calculated from the theoretical distribution, minimizes reliance on the information from doctor’s scoring. We use the manual score only to identify the characteristics which are useful for the sleep stage separation.

The composition of the paper is as follows:

In Section 2 we define various characteristics and methods used in our analysis of neonate sleep EEG patterns. We begin with the spectral characteristics and some nonlinear measures in Section 2.1 and move on to the fractional dimension in Section 2.2.

Section 3 analyzes concrete recordings for 21 fullterm and 16 preterm healthy neonates of 40 week post-conceptional age each. In Section 3.1 we use the non-parametric Wilcoxon test to select characteristics that are best suited for the separation of quiet and active sleep stages for both fullterm and preterm populations. A nonlinear transformation of some diagnostic sequences is utilized to guarantee the symmetry of the data distribution and thus the applicability of the Wilcoxon test. In Section 3.2 optimal characteristics to discriminate between fullterm and preterm infants within the same sleep stage are identified.

Finally, Section 4 develops and tests an automatic procedure for the separation of neonatal sleep stages using EEG-based change-point detection algorithms and diagnostic sequences that incorporate the significant characteristics identified in Section 3.