Sleep spindle detection through amplitude–frequency normal modelling

Abstract

Manual scoring of sleep spindles can be very time-consuming, and achieving accurate manual scoring on a long-term recording requires high and sustained levels of vigilance, which makes it a highly demanding task with the associated risk of decreased diagnosis accuracy. Although automatic spindle detection would be attractive, most available algorithms are sensitive to variations in spindle amplitude and frequency that occur between both subjects and derivations, reducing their effectiveness.

We propose here an algorithm that models the amplitude–frequency spindle distribution with a bivariate normal distribution (one normal model per derivation). Subsequently, spindles are detected when their amplitude–frequency characteristics are included within a given tolerance interval of the corresponding model. As a consequence, spindle detection is not directly based on amplitude and frequency thresholds, but instead on a spindle distribution model that is automatically adapted to each individual subject and derivation.

The algorithm was first assessed against the scoring of one sleep scoring expert on EEG samples from seven healthy children. Afterward, a second study compared performance of two additional experts versus the algorithm on a dataset of six EEG samples from adult patients suffering from different pathologies, to submit the method to more challenging and clinically realistic conditions. Smaller and shorter spindles were more difficult to evaluate, as false positives and false negatives showed lower amplitude and smaller length than true positives. In both studies, normal modelling enhanced performance compared to fixed amplitude and frequency thresholds. Normal modelling is therefore attractive, as it enhances spindle detection quality.

Introduction

Sleep spindles’ detection is of major importance for staging sleep as well as in the field of sleep research. A spindle is commonly defined as a group of rhythmic waves characterized by progressively increasing, then gradually decreasing amplitude, that may be present in low voltage background EEG, superimposed to delta activity, or temporally locked to a vertex sharp wave and to a K complex (De Gennaro and Ferrara, 2003). Spindles are one of the hallmarks of Non-Rapid Eye Movement (NREM) stage 2 sleep, both in adults and children (Iber et al., 2007, Grigg-Damberger et al., 2007). They are affected by normal ageing, by pathological ageing (De Gennaro and Ferrara, 2003, Petit et al., 2004), as well as by brain pathology; therefore, the spindles could be used as a marker of normal brain functional development (Fogel and Smith, 2011, Petit et al., 2004). Furthermore, they have been proposed as ideal candidates to induce, in the neocortex, long-term synaptic changes which are necessary for learning and memory (Destexhe and Sejnowski, 2001, Peigneux and Smith, 2010), even after a daytime nap (Schmidt et al., 2006). Accordingly, recent evidence in adults suggests that spindles are highly correlated with intellectual ability (IQ tests) and with an overnight improvement in performance after learning new material (Fogel and Smith, 2011).

Sleep spindles rapidly develop during the first three months of age. Sizeable maturational changes can be observed throughout the developmental phase in terms of frequency, amplitude, and amount. These changes reputedly reflect the development of thalamocortical structures and the maturation of the physiological system that produces spindles (De Gennaro and Ferrara, 2003), indicating that spindles are longitudinal markers for the ontogenic evolution of brain functioning. Conversely, modifications in sleep and spindle variables are observed with age (De Gennaro and Ferrara, 2003). Sleep spindles may thus represent, at the electrophysiological level, an ideal mechanism that reflects long-term synaptic changes in the neocortex (Fogel and Smith, 2011). Therefore, modulations and alterations in sleep spindle activity should be explored in childhood developmental disorders with associated cognitive impairments and in brain pathologies (Reeves and Klass, 1998, Shibagaki and Kiyono, 1983), that could also be associated with impaired sleep-dependent consolidation processes (Urbain et al., 2011, Van Bogaert et al., 2012, Chan et al., 2011).

Manual scoring of spindles is time-consuming for recordings that typically show 1000 spindles (Acir and Güzeliş, 2004). Achieving accurate manual scoring on long-term recordings requires a high level of vigilance, resulting in a highly demanding task that augments the risk of decreased accuracy in the diagnosis, especially for sleep-related studies, for which precise information (such as spindle's amplitude, frequency, and length) is often required. Beside the laborious aspect of the task, visual analysis involves some subjectivity (inter-human agreement is estimated to be around 80–90% (Campbell et al., 1980), and degree of consent is 70 ± 8% (Zygierewicz et al., 1999)). A reliable spindle detection is therefore attractive, as it would enhance the speed, accuracy, and inter-rater agreement of spindle scoring.

Various spindle detection algorithms have been previously proposed. Recent ones are based on methods that include fuzzy logic (Huupponen et al., 2000a, Huupponen et al., 2003), neural network (Shimada et al., 2000, Huupponen et al., 2000b, Acir and Güzeliş, 2004, Ventouras et al., 2005, Güneş et al., 2011), bandpass filter (Clemens et al., 2005, Huupponen et al., 2007), fast time frequency transform (Knoblauch et al., 2003a, Knoblauch et al., 2003b), Fourier transform (Huupponen et al., 2007, Duman et al., 2009), wavelet transform (Duman et al., 2009), Gabor transform (Schönwald et al., 2003) and matching pursuit (Durka et al., 2005, Schönwald et al., 2006, Ktonas et al., 2009).

Fewer studies have investigated spindle detection in children (Grigg-Damberger et al., 2007, Causa et al., 2010). However, assessing automated spindles detection, specifically in childhood, seems important as spindles show modification in amplitude, frequency, length, density, interspindle interval and topological distribution from infancy to adolescence (Nagata et al., 1996, Shinomiya et al., 1999, Scholle et al., 2007); therefore, automated algorithms should be able to adapt to those variations. Published spindle detection algorithms in children and in infants are based on empirical-mode decomposition and Hilbert–Huang transform (Causa et al., 2010), expert procedure and fuzzy logic (Held et al., 2004, Held et al., 2006), peak identification (Estévez et al., 2002), merge neural gas model (Estévez et al., 2007), and neuro fuzzy approach (Heiss et al., 2002).

The majority of the proposed algorithms are – directly or indirectly – based on amplitude–frequency analysis, thus banking on spindle definition and mimicking visual analysis. One of the major difficulties encountered with these detection methods is the setting of suitable thresholds for the amplitude and the frequency. Spindle frequency is traditionally defined as 12–14 Hz (Rechtschaffen and Kales, 1968, Grigg-Damberger et al., 2007), but may often extend to both higher and lower frequencies. Therefore, the ‘classical’ 12–14 Hz spindle definition is believed to be too narrow (Jankel and Niedermeyer, 1985). The difficulty in finding the optimum frequency bounds has produced a large number of proposed values, among them: 11.5–15 Hz (Fish et al., 1988), 11.5–16 Hz (Zeitlhofer et al., 1997), 11–15 Hz (Ktonas et al., 2009), 11–16 Hz (Clemens et al., 2005), 10.5–16 Hz (Ventouras et al., 2005, Huupponen et al., 2007), and 10–16 Hz (Zygierewicz et al., 1999, Huupponen et al., 2000a, Estévez et al., 2002). Beside the often cited 12–14 Hz frequency range proposed by the National Institute of Neurological Diseases and Blindness of the U.S. Department of Health, Education, and Welfare (Rechtschaffen and Kales, 1968), various organizations have suggested other values to score spindles: 11–16 Hz by the American Academy of Sleep Medicine (Iber et al., 2007), 11–15 Hz by the International Federation of Clinical Neurophysiology (Noachtar et al., 1999), and 12–16 Hz by the Japanese Society of Sleep Research (Hori et al., 2001). Finally, the Paediatric Task Force published a visual scoring of sleep and arousal in infants and children (Grigg-Damberger et al., 2007), but did not propose a definite spindle frequency range, underlying the variety of criteria found in the literature. Similarly, the spindle's minimum amplitude is difficult to determine, independently of the chosen definition of the amplitude (Fish et al., 1988). Previous researchers have set arbitrary lower amplitude thresholds at 8, 10, 12, and 14 μV (Fish et al., 1988), 10 μV (Estévez et al., 2002, Ventouras et al., 2005), 15 μV (Zygierewicz et al., 1999), and 25 μV (Zeitlhofer et al., 1997).

Universal amplitude and frequency intervals are difficult to define for spindles because of the large variability across subjects and derivations. Indeed, there is a large variability in both amplitude (Huupponen et al., 2000a, Clemens et al., 2005) and frequency (Zeitlhofer et al., 1997) across subjects and in both amplitude and frequency across derivations (Zeitlhofer et al., 1997). Therefore, it seems difficult to design a spindle detection algorithm without adapting to this variability.

Consequently, we propose to consider that both amplitude and frequency ranges vary between subjects and derivations, and to tailor thresholds accordingly. We designed an algorithm that models the amplitude–frequency spindle distribution with a bivariate normal distribution (one normal model per derivation). Subsequently, spindles are detected when their amplitude–frequency characteristics are included within a given tolerance interval (TI) of the corresponding model. As a consequence, spindle detection is not directly based on amplitude and frequency thresholds, but instead on a spindle distribution model that is automatically adapted to each individual and each derivation. Furthermore, the TI may be adjusted for more sensitive or selective detection.

Taking this idea one step further, spindle detection may be considered probabilistic in nature. In this regard, a given spindle has a probability of detection, which may be high or low depending on its amplitude, frequency, morphology, background activity, and other attributes. An identical approach has previously been proposed for EEG spikes (Wilson et al., 1996). For instance, an event with a shape similar to a spindle, but with a low amplitude (compared to its background) or with a particularly high/low frequency, could be accepted as a spindle by some specialists and not by others, whereas a clear spindle would be accepted by all specialists. Our approach could also be seen as a probabilistic estimation of spindle events as a function of their amplitudes and frequencies.

To evaluate the ability of our algorithm to detect spindles, we have analyzed two data sets. The algorithm was first trained and tested against the scoring of one expert on EEG samples that we obtained in seven healthy children (Study 1). The objective was to evaluate the ability of the algorithm to detect spindles for this category. Afterward, a second study compared performance of two additional experts versus the algorithm on six EEG samples from adult patients suffering from different pathologies (Study 2). The purpose was to submit the method to more challenging conditions, since pathologies may affect the shape of the spindles. Recording conditions in this database were suboptimal but closer to common clinical situations (i.e. only three derivations were available, sampling frequency was lower, artefacts were more frequent, and noise levels were higher).