A comparison between detectors of high frequency oscillations
Highlights
► The automatic detection of HFOs is crucial to propel the clinical use of HFOs as biomarkers of epileptogenic tissue. ► A comparison of existing detectors on the same dataset is presented to analyze their performance and to emphasize the issues involved in validation. ► Optimizing on a particular type of data could improve performance in any detector.
Introduction
High frequency oscillations (HFOs) are emerging as biomarkers of epileptogenic tissue that could help in the identification of epileptic or potentially epileptic regions during intracranial investigations. This could help the delineation of the surgical extent and in the prediction of surgical outcome (Jacobs et al., 2010).
HFOs are spontaneous EEG patterns in the range of 80–500 Hz, consisting of at least four oscillations that can be “clearly” distinguished from background. HFOs were first recorded with microelectrodes (20–40 μm in diameter) implanted in temporal regions (Bragin et al., 1999a), and recently with clinical macro-electrodes in temporal and neocortical regions (Jirsch et al., 2006, Urrestarazu et al., 2007, Worrell et al., 2008). When recorded with macro-electrodes, HFOs are characterized by a typical duration of 30–100 ms, an inter-event interval of at least 25 ms, and an amplitude of 10–100 μV. These EEG patterns occurring in the absence of specific stimuli, have been recorded during interictal (Staba et al., 2002, Urrestarazu et al., 2007), pre-ictal (Jacobs et al., 2009) and ictal (Jirsch et al., 2006) periods. Interictally, they can be identified more frequently during slow wave sleep than during wakefulness (Staba et al., 2004, Bagshaw et al., 2009).
During interictal periods, higher rates of HFOs were observed in the seizure onset zone (SOZ) than in other areas (Bragin et al., 1999b, Urrestarazu et al., 2007). Even though a large proportion of HFOs co-occur with spikes, HFOs can occur also in non-spiking channels or independently from spikes (Jacobs et al., 2008). The ranking of channels according to rate (Zelmann et al., 2009) indicated that HFOs remained confined to the same region during interictal and ictal periods, while spikes presented a wider spread during seizures than interictally (Zijlmans et al., 2011). Moreover, a postsurgical study showed a correlation between surgical outcome and removal of channels with high HFO rates (Jacobs et al., 2010). In summary, interictal HFOs seem to be a reliable biomarker of tissue capable of producing seizures.
Visual marking of HFOs provided a good understanding of the relation of HFOs with epilepsy (Urrestarazu et al., 2007, Zijlmans et al., 2009, Jacobs et al., 2010). However, visual marking is highly time consuming (it takes about 10 h to visually mark HFOs in a 10-channel 10-min recording) and subjectivity is inevitable. Thus, the development of automatic HFOs detectors is crucial for the systematic study of HFOs and for their eventual clinical application. The lack of a formal definition makes the detection of HFOs difficult and subjective.
Only a handful of automatic detectors based on different energy functions exist (Staba et al., 2002, Gardner et al., 2007, Crepon et al., 2010, Zelmann et al., 2010). In the first three the energy threshold is computed based on the segment of EEG under consideration, under the assumption that HFOs are rare events. On the contrary, the detector developed at the Montreal Neurological Institute (MNI) first detects baseline segments and then uses this information to compute the local energy threshold (Zelmann et al., 2010). Each detector was developed for a different frequency band (above 200 Hz, (Crepon et al., 2010), gamma band (Gardner et al., 2007)); a particular location (only temporal (Staba et al., 2002)); or electrode size (microelectrodes in (Staba et al., 2002), micro- and clinical macro-electrodes in (Worrell et al., 2008)), small clinical macro-electrodes (Zelmann et al., 2010). It is therefore important to test all the detectors on the same dataset.
In this study, we present the complete version of the MNI detector and provide a comparison between these four detectors on the same data. The new MNI detector includes a feature to deal separately with channels with very frequent or constant HF rhythmic activity. We hypothesize that all the detectors will behave similarly in those channels where HFOs are rare events, but that the MNI detector will outperform the others in channels with very frequent HFOs.
Section snippets
Patient information
Between September 2004 and April 2008, 45 patients with medically intractable epilepsy underwent depth macro-electrode (surface area 0.8 mm2) implantation at the Montreal Neurological Hospital. The depth EEG (SEEG), recorded with the Harmonie system (Stellate, Montreal, Canada), was low-pass filtered at 500 Hz and sampled at 2000 Hz, allowing for the identification of HFOs. Twenty patients were randomly selected, but one had to be excluded due to continuous artefacts. All patients gave informed
Original configurations
We compared our detector with the others considering the parameters, filter and thresholds in the original publications (see Table 1). Fig. 3 shows the ROC curves across channels. The MNI detector considers differently channels with more than 5 s/min of detected baselines (170 channels) and channels with less than 5 s/min of baseline (127 channels). Results are presented separately for the two groups. With the original configuration, the AUC for “Channels with Baseline” was: 0.994 for the RMS
Discussion
Visual markings of HFOs provided a good understanding of the relation of HFOs with epilepsy (Urrestarazu et al., 2007, Jacobs et al., 2008, Jacobs et al., 2010, Zijlmans et al., 2009). But this manual procedure is highly time consuming and subjective. For the systematic study of HFOs and for a future clinical application, robust automatic detection of HFOs is necessary. In this study, we presented a comparison of performance of HFO detectors based on energy on the same dataset.
In channels with
Acknowledgements
This work was supported by Grants MOP-10189 and MOP-102710 from the Canadian Institutes of Health Research. RZ was supported by National Science and Engineering Research Council (NSERC) Postgraduate Scholarship (PGSD).
References (29)
- et al.
Pitfalls of high-pass filtering for detecting epileptic oscillations: a technical note on “false” ripples
Clin Neurophysiol
(2010) - et al.
Human and automated detection of high-frequency oscillations in clinical intracranial EEG recordings
Clin Neurophysiol
(2007) - et al.
State dependent spike detection: validation
Electroencephalogr Clin Neurophysiol
(1992) - et al.
Wavelet entropy: a new tool for analysis of short duration brain electrical signals
J Neurosci Methods
(2001) - et al.
Improving the identification of high frequency oscillations
Clin Neurophysiol
(2009) - et al.
Ictal and Interictal high frequency oscillations in patients with focal epilepsy
Clin Neurophysiol
(2011) - et al.
Effect of sleep stage on interictal high-frequency oscillations recorded from depth macroelectrodes in patients with focal epilepsy
Epilepsia
(2009) - et al.
Unsupervised classification of high-frequency oscillations in human neocortical epilepsy and control patients
J Neurophysiol
(2010) - et al.
Hippocampal and entorhinal cortex high-frequency oscillations (100–500 Hz) in human epileptic brain and in kainic acid-treated rats with chronic seizures
Epilepsia
(1999) - et al.
Electrophysiologic analysis of a chronic seizure model after unilateral hippocampal KA injection
Epilepsia
(1999)