A general automatic method for the analysis of NREM sleep microstructure

Abstract

Objective

To define a unified method for the automatic recognition and quantitative description of EEG phasic events of sleep microstructure occurring during NREM sleep, particularly arousals, phase A subtypes of cyclic alternating pattern and spindles.

Methods

The NREM sleep EEG of 10 normal young subjects was examined in order to recognize formal phasic events of sleep microstructure. The following ‘formal’ events (i.e. events defined exclusively on the basis of automatic analysis criteria) were classified: arousals, A1-phases (A-phases not including arousals) and A2- and A3-phases (A-phases including arousals). Spindle bursts, corresponding to visually recognized spindles, were also formally defined. The identification of these events was carried out following a three-step procedure: (1) computation of band-related descriptors derived from the EEG signal, (2) introduction of suitable thresholds and (3) application of simple logical principles, i.e. an exclusion principle and an overlapping principle.

Results

Formal A-phases, arousals and spindle bursts showed spectral characteristics which were consistent with visual inspection. The value of the parameter Correctness for the recognition of the A-phases was 83.5%. In particular, the different physiological distribution of the A-phases in Stage 2 preceding slow wave sleep with respect to Stage 2 preceding REM sleep was confirmed.

Conclusions

The proposed method provides a unified quantitative approach to the study of sleep microstructure. Visually defined events can be reliably identified by means of automatic recognition.

Introduction

Transient EEG waveforms or, more generally, transient polygraphic events play a twofold role in sleep analysis.

First, their occurrence and their rate of occurrence are fundamental in visual scoring of polysomnographic measures according to Rechtschaffen and Kales' [1] rules. Events such as K-complexes, alpha bursts, spindles and vertex sharp waves are well-known markers for stage scoring.

Many transient events are not only landmarks for sleep stage recognition but their occurrence plays an active role in the dynamics of the sleep profile. This consideration is supported by the increasing interest in arousals and other microstructure components of sleep. The term microstructure refers to EEG features below the time dimension of the conventional 20–30 s scoring epoch [1]. The recognition of microstructure events provides physiological and clinical information which integrates the macrostructure measures obtained with the conventional staging system.

The criteria for recognition and classification of arousals were established in 1992 by the American Sleep Disorders Association [2]. Agreement was found on the following definition: ‘An EEG arousal is an abrupt shift in EEG frequency, which may include theta, alpha and/or frequencies greater than 13 Hz, but not spindles. The EEG frequency shift must be 3 s or greater in duration to be scored as an arousal.’

Another significant microstructure phenomenon is cyclic alternating pattern (CAP) [3]. Criteria for the recognition and classification of CAP were given in 2001 in a Consensus Report [4], defined as the following: ‘The CAP is a periodic EEG activity of non-REM sleep. CAP is characterized by sequences of transient electrocortical events that are distinct from background EEG activity and recur at up to 1 min intervals.’ Recurring electrocortical events are called A-phases and are divided into three classes: A1, A2, and A3. CAP appears spontaneously but also in association with identifiable sleep pathophysiologies such as sleep disordered breathing [5] and periodic limb movement disorder [6]. CAP evaluation includes the periodicity dimension in the arousal process and attributes different levels of cerebral activation as expressed by the phase A subtypes.

The phase A1 subtypes are composed of K-complexes and delta bursts; they prevail in Stage 2 that precedes slow wave sleep. Phases A2 and A3 of CAP are composed of mixed EEG patterns including both slow and rapid activities; they prevail in Stage 2 that precedes REM sleep. An extensive overlap between ASDA arousals and subtypes A2 and A3 has been demonstrated [7]. Due to their complex EEG morphologies, arousals and CAP include a variety of transient changes in different frequency bands.

A number of automatic and quantitative methods for the analysis of microstructure phenomena have been carried out in the last decades. These methods often present high agreement with visual analysis. We limit ourselves to reporting briefly some of the most interesting events. A general model for the analysis of sleep spindles and alpha rhythms was proposed in the 1980s [8], [9]. It was based on the idea that EEG phenomena are generated by excitatory and inhibitory neuronal populations interacting by means of feedback loops. This model was then also applied to the detection of vertex waves and K complexes [10]. In 1994, Jobert et al. [11] suggested the application of the Wavelet Transform to the analysis of transient events and supported this idea with preliminary significant results. Pardey et al. [12] proposed a neural network EEG analysis system based on an autoregressive modeling of the signal; the EEG was quantified on a continuous scale which was not linearly related to conventional sleep stages. McKeown et al. [13] described a method for detecting stage changes in the EEG, which was based on the properties of a dimensionless function calculated by using independent component analysis. De Carli et al. [14] detected arousals applying the Wavelet Transform to two bipolar EEG traces and one EMG derivation. In a following study the same group [15] compared the mean power values of the entire arousal with the immediately preceding 3.5 s and found an enhancement in the band of delta power relative to background. An automatic system for the detection of CAP sequences was proposed by Rosa et al. [16]. The system consisted of three parts: a model-based maximum likelihood estimator, a variable length template-matched filter, and a state machine rule-based decision subsystem. The Matching Pursuit Procedure, based on the decomposition of the signal into waveforms with good localization in time and frequency, was applied to the identification and parameterization of spindles [17], [18]. The method for spindle detection described by Huupponen et al. [19] was based on the application of a variable threshold the value of which was estimated by Bayesian analysis. More recently, Huupponen et al. [20] identified, via a mean frequency measure and FFT, sleep oscillations with period times of 50–150 s having a relatively large amplitude.

There are two different approaches for assessment of automatic analysis of sleep EEG. One is the agreement with the output of visual analysis. A remarkable example is provided by the so-called ‘hybrid’ systems, developed in the 1970s, which are in part analog and in part digital. These systems achieve a high rate of agreement with visual scoring [21], [22]. A second approach is based on the idea that automatic signal processing can provide additional information with respect to that given by visual analysis. An example can be given by the characteristic damped-oscillation pattern of the delta rhythm [23], [24], [25], which provides important information, the details of which are not contained in the histogram.

According to Kubicki et al. [26], an automatic analysis of sleep EEG should emphasize the strengths of the computer, with a substantial independence from visual analysis. The Rechtschaffen and Kales' rules [1] are conventional criteria but can be inadequate for computer-based automatic analysis.

If these considerations can be applied to the analysis of sleep profile, they are even more suitable for the analysis of phasic events of sleep microstructure. The aim of the study was to define a unified method that stems from the characteristics of visual analysis and introduces new criteria closely connected to the discriminating properties of automatic analysis.