Abstracts

Rationale: High frequency oscillations (HFOs) are considered as a candidate biomarker of tissue that can generate epileptic seizures. Filtering the EEG signal is a common way to detect them, however some other broad-band EEG components (epileptic spikes, sharp waves and artifacts) may disturb the detection process by generating false oscillations in the filtered signal. We present a method allowing the separation of sharp transients with HFOs from those without HFOs by detecting the oscillations in the raw, unfiltered signal. Methods: A classification method based on temporal features is proposed to differentiate real oscillations from false ones (those generated by filtering effect). The presence of an isolated peak between 80-500 Hz in the time-frequency plane was considered as the definition of real HFOs. As an example, two events are shown in figure 1. The spike at the left shows no isolated peak in the ripple band (80-250Hz), while in the right, a peak appears there. Both events are visually marked as HFOs by experts. The proposed method was evaluated on one-minute intracranial EEG of 15 patients. The data of 5 patients was selected randomly as the training set and the other 10 were used as the test set. In total 48023 sharp events, 2856 ripples and 1450 fast ripples (FR, 250-500 Hz) were marked. 1343 of the ripples and 921 of the FRs were coinciding with sharp events, and were selected for the investigation. Several features including the number of oscillations in the raw signal, the durations of HFOs and of the sharp events, and the amplitude of raw and filtered signal were extracted for each event. Forward feature selection method was then used to derive significant features quantifying the discrimination between real and false oscillations. The features in the ripple and FR band yielding the best classification performance were selected for the testing procedure. A supervised classifier (support vector machine) was then trained in the feature space. Its performance was evaluated using cross validation applied to the testing data. Results: Results show that around 50% of visually marked HFOs co-occurring with sharp transients are not real HFOs. It is also observed that the number of oscillations of real HFOs is significantly higher than that of false ones (p<0.0001). Moreover, the ratio of raw signal amplitude to the filtered signal amplitude is higher among the sharp events concurring with real HFOs than falsely detected HFOs (p<0.001, bottom row of figure 1). The proposed method is capable of separating sharp transients with real ripples from those generating false ripples with the mean accuracy of 80% and mean sensitivity of 71%. In FR band, the mean accuracy is 82%, and the mean sensitivity is 79%. Conclusions: We demonstrate that a good proportion of visually-marked HFOs occurring at the same time as spikes do not correspond to any oscillations visible on the unfiltered signal. We also conclude that some spikes have high-frequency components, while others are restricted to lower frequency bands. They may have different roles in epileptogenicity.