Instantaneous frequency based newborn EEG seizure characterisation

The electroencephalogram (EEG), used to noninvasively monitor brain activity, remains the most reliable tool in the diagnosis of neonatal seizures. Due to their nonstationary and multi-component nature, newborn EEG seizures are better represented in the joint time–frequency domain than in either the time domain or the frequency domain. Characterising newborn EEG seizure nonstationarities helps to better understand their time-varying nature and, therefore, allow developing efficient signal processing methods for both modelling and seizure detection and classification. In this article, we used the instantaneous frequency (IF) extracted from a time–frequency distribution to characterise newborn EEG seizures. We fitted four frequency modulated (FM) models to the extracted IFs, namely a linear FM, a piecewise-linear FM, a sinusoidal FM, and a hyperbolic FM. Using a database of 30-s EEG seizure epochs acquired from 35 newborns, we were able to show that, depending on EEG channel, the sinusoidal and piecewise-linear FM models best fitted 80–98% of seizure epochs. To further characterise the EEG seizures, we calculated the mean frequency and frequency span of the extracted IFs. We showed that in the majority of the cases (>95%), the mean frequency resides in the 0.6–3 Hz band with a frequency span of 0.2–1 Hz. In terms of the frequency of occurrence of the four seizure models, the statistical analysis showed that there is no significant difference (p = 0.332) between the two hemispheres. The results also indicate that there is no significant differences between the two hemispheres in terms of the mean frequency (p = 0.186) and the frequency span (p = 0.302). Introduction Seizures tend to happen more frequently in the neonatal period than at any other stage in life [1]. The reported incidence of seizure ranges from 1 to 3 per 1 000 live births in term infants and 10 to 15 per 1 000 live births in preterm infants [2]. Seizures usually arise as the result of excessive, synchronous electrical discharge, of neurons within the central nervous system [3,4]. Although not a disease in themselves, seizures are the most prominent manifestation of neurological dysfunction in the newborn [4]. They often suggest underlying disease processes which may cause irreversible damage to the developing neonatal brains and have demonstrated association with infant mortality and long term morbidity [5]. The underlying brain conditions associated with seizures in the neonates include hypoxic-ischemic encephalopathy, brain *Correspondence: [email protected] 1University of Queensland Centre for Clinical Research, The University of Queensland, Brisbane, QLD 4029, Australia 2School of Mechanical and Chemical Engineering, The University of Western Australia, 35 Stirling Highway, Crawley, Perth, WA 6009, Australia Full list of author information is available at the end of the article haemorrhage, stroke and meningitis [6,7]. It is therefore critical to recognise neonatal seizures in their early stages to allow timely medical intervention. Clinical assessment of seizures in the neonates is difficult and unreliable as many neonatal seizures occur either in the absence of any clinical signs or accompanied by only subtle ones [8,9]. The clinical diagnosis is further hampered by the frequent administration of sedative or paralytic agents to the newborn patients in neonatal intensive care units. Recorded via electrodes attached to the scalp, electroencephalogram (EEG) noninvasively measures the electrical activities of the brain and provides useful information about its state. In the neonates, EEG is often the first test to reveal clinically unsuspected seizures and remains the only reliable method for the identification and diagnosis of subclinical seizures [10]. Besides being an effective tool for diagnosis, newborn EEG also correlates with longterm neurodevelopment outcome [11]. EEG analysis can also assist in the design of automated methods for seizure detection, classification, and source localisation among other numerous applications [12]. © 2012 Mesbah et al.; licensee Springer. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Mesbah et al. EURASIP Journal on Advances in Signal Processing 2012, 2012:143 Page 2 of 11 http://asp.eurasipjournals.com/content/2012/1/143 Seizures appear in EEG as sudden, repetitive, evolving, stereotyped waveforms that last at least 10 s and have a definite beginning, middle, and end [13]. Their frequencycontent varies with time [14-16] and is, therefore, best represented in the joint time–frequency domain instead of either the time domain or the frequency domain. Time– frequency methods, such as quadratic time–frequency distributions (TFDs) [17], facilitate the analysis of EEG seizure signals by exploiting their spectral energy variation with time. Electroencephalogram patterns in neonatal seizures are highly variable with complex and varied morphology, frequency, and topography [18]. In this study, we analysed newborn EEGs using a quadratic TFD with separable kernel to characterise seizures in the time–frequency domain. We extracted the instantaneous frequency (IF) from the TFD to determine its modulation law; an important descriptive characteristic of EEG seizure [1416,19,20]. The IFs were extracted using amethod designed specifically for multi-component signals [21] and applied to a separable kernel TFD with optimised parameters. These extracted IFs were fitted to the four frequencymodulated (FM) models previously linked with newborn EEG seizure [14,21]. To further characterise the extracted IFs, we computed their mean frequency and frequency span. Characterisation of newborn EEG seizures is an essential step in a number of applications such as modelling [16,22] and detection/classification [15,19,20]. Time–frequency signal processing Time–frequency signal processing arose due to the need for accurate representation and efficient analysis and processing of nonstationary signals [17]. Nonstationary signals are very common natural phenomena. They are characterised by their time-varying frequency content which make them unsuitable for analysis by traditional methods, such as Fourier transform, that assume stationarity. Time– frequency signal analysis uses TFDs to represent signals in the joint time–frequency domain and is, therefore, capable of tracking signals’ spectral change over time. Quadratic time–frequency distributions Quadratic TFDs have been extensively used in the analysis and processing of nonstationary signals in a number of practical applications. They can be mathematically formulated as [17]: ρz(t, f ) = Wz(t, f ) ∗∗ (t,f ) γ (t, f ) (1) where ρz(t, f ) denotes the TFD,Wz(t, f ) the Wigner–Ville distribution (WVD), γ (t, f ) the time–frequency kernel, and ∗ the linear convolution operation. The above formulation can also be given in any of the other three joint domains [time-lag (t, τ), Doppler-frequency (ν, f ), and Doppler-lag (ν, τ)] that are linked to the time–frequency domain via the Fourier transform [17]. With a time– frequency kernel defined by γ (t, f ) = δ(t)δ(f ), where δ is the Dirac delta function, theWVD is the most basic member of the quadratic class. For a real-valued signal s(t), the WVD is defined as [17,23]: