Seizure detection algorithm for neonates based on wave-sequence analysis

Abstract

Objective

The description and evaluation of the performance of a new real-time seizure detection algorithm in the newborn infant.

Methods

The algorithm includes parallel fragmentation of EEG signal into waves; wave-feature extraction and averaging; elementary, preliminary and final detection. The algorithm detects EEG waves with heightened regularity, using wave intervals, amplitudes and shapes. The performance of the algorithm was assessed with the use of event-based and liberal and conservative time-based approaches and compared with the performance of Gotman's and Liu's algorithms.

Results

The algorithm was assessed on multi-channel EEG records of 55 neonates including 17 with seizures. The algorithm showed sensitivities ranging 83–95% with positive predictive values (PPV) 48–77%. There were 2.0 false positive detections per hour. In comparison, Gotman's algorithm (with 30 s gap-closing procedure) displayed sensitivities of 45–88% and PPV 29–56%; with 7.4 false positives per hour and Liu's algorithm displayed sensitivities of 96–99%, and PPV 10–25%; with 15.7 false positives per hour.

Conclusions

The wave-sequence analysis based algorithm displayed higher sensitivity, higher PPV and a substantially lower level of false positives than two previously published algorithms.

Significance

The proposed algorithm provides a basis for major improvements in neonatal seizure detection and monitoring.

Introduction

Seizures are a common sign of neurological disorder in neonates (Lombroso, 1996, Volpe, 2001), and can be caused by virtually any condition that affects neonatal brain function (Tharp, 2002). The occurrence of seizures in the newborn are associated with increased morbidity and mortality (Scher et al., 1989, Legido et al., 1991, Lombroso, 1996). The incidence of seizures is greater in the neonatal period than at any other time in life (Mizrahi, 1999, Patrizi et al., 2003), occurring in up to 0.5% of newborn infants (Tharp, 2002). However, there are major clinical challenges in the recognition of neonatal seizures, related partly to the high incidence of sub-clinical seizures (55–85%) and in that sick newborns may be pharmacologically paralysed in the intensive care setting (Hellstrőm-Westas et al., 1985, Bye and Flanagan, 1995, Gotman et al., 1997a).

Neonatal seizures have extremely variable morphology, frequency, and topography, even within the same patient, and are much more variable than seizures in adults (Shewmon, 1990, Lombroso, 1996, Patrizi et al., 2003). Due to the challenges in 24 h access to the gold standard of conventional EEG, many neonatal units now utilise amplitude-integrated EEG (aEEG) to monitor brain functions and assist in seizure recognition (e.g. Murdoch-Eaton et al., 2001, Procaccio et al., 2001, Hellstrőm-Westas et al., 2003). The amplitude-integrated EEG has historically been produced from a single EEG-channel that has been bandpass filtered between 2 and 15 Hz, with amplification of the higher frequency components, rectified, then submitted to a part-linear, part-logarithmic amplitude compression, ‘peak-smoothed’ and time-compressed (Maynard et al., 1969, Prior et al., 1973). Patterns emerging on the aEEG trace are used to classify background EEG activity and to identify seizures (e.g. Greisen and Tape-recorded, 1994, Hellstrőm-Westas et al., 2003, de Vries and Hellstrőm-Westas, 2005). However, it has been recognized that non-expert use of aEEG traces can be difficult without access to the raw EEG record (Thordstein et al., 2000), and that focal, low amplitude or shorter than 1 min duration seizures are often missed (Rennie et al., 2004). Attempts to automatically detect seizures are using the aEEG trace alone have not been made until recently (Lommen et al., 2005), and seizure length <1 min cannot be detected by this method. Another approach to seizure detection is based on visual assessment of the compressed spectral analysis graph of EEG recorded continuously at the cotside (Aziz et al., 1986).

Several automatic seizure detection methods for conventional EEG in neonates have been developed, though not many have been systematically tested. Liu et al. (1992) suggested looking for heightened regularity in the positions of the first 4 peaks in the autocorrelation function. Nagasubramanian et al. (1997) identified heightened regularity by a decrease in the spectral exponent and decreases in the variability of the power in frequency bands, approximately corresponding to ‘clinically important’ frequencies (1–15 Hz). They proposed a wavelet based calculation to efficiently perform this identification. Roessgen et al. (1998) suggested a modification of da Silva's ‘local EEG model’ (da Silva et al., 1974) with an additional saw-tooth input, representing seizure activity. Using Maximum Likelihood techniques to estimate model parameters allows the power spectrum of EEG to be described as the sum of the background power spectrum, seizure power spectrum and Gaussian noise. Seizure activity is detected when the ratio of power in the modelled seizure to background spectrum exceeds a threshold value in favour of seizure spectrum. Celka and Colditz, 2002a, Celka and Colditz, 2002b presented an alternative to Roessgen et al. (1998) model of EEG. The background activity is modelled as a non-linear transformation of a Gaussian ARMA-process which has an ‘exact’ match to the observed distribution of background EEG voltages. An EEG signal is, therefore, transformed by applying the inverse of the non-linear transformation and ARMA-filter. Finally, singular spectrum analysis is used to test if the output looks like Gaussian white noise, with a false result indicating seizure. Smit et al. (2004) used the synchronization likelihood to measure statistical interdependence between two time series, i.e. between a pair of EEG channels, assuming that during seizure activity the synchronization likelihood would rise. Boashash and Mesbah, 2001, Hassanpour et al., 2004 presented some promising findings employing joint time–frequency analysis.

The only commercially available and widely tested algorithm for neonatal seizure detection has been developed by Jean Gotman (Gotman et al., 1997a). Three methods are used with a 10 s sliding foreground window (moving by 2.5 s increments); which is compared with a preceding 20 s background window separated from it by a 1 min interval. The first method is intended for the detection of rhythmic discharges of 0.5–10 Hz and is principally based on spectral analysis (Qu and Gotman, 1997). An increase in the power of the dominant power spectrum peak indicates the presence of seizure, subject to a range of conditions (Gotman et al., 1997a). The other methods are based on half-wave decomposition of EEG signal (Gotman and Gloor, 1976). The second method utilises the detection of multiple spikes (Gotman and Gloor, 1976, Gotman et al., 1979), where half-waves of certain length, sharpness and relative amplitude are classified as spikes, and the presence of 6 spikes in the foreground window indicates a seizure (Gotman et al., 1997a). The third method is designed to detect very slow rhythmic discharges. EEG is low-pass filtered with a cut-off frequency of 1 Hz (Gotman et al., 1997a). Then, to qualify as a seizure the foreground window has to exceed thresholds for relative amplitude and coefficient of correlation of half-wave durations (Gotman, 1982, Gotman, 1990).

All algorithms described above have one feature in common. They apply one or other method to a fixed time interval (usually a sliding window) in the EEG trace, trying to extract interval features and compare them with background features. The aim of this article is to present and evaluate a real-time algorithm based on the analysis not of time intervals, but of the wave sequences in the EEG recording.