Automated Detection of Fast Ripples and Rejection of Artifacts in Human Scalp Electroencephalogram
High frequency oscillations (HFOs) are a promising biomarker of epileptic tissues. Higher rates of HFOs are observed in seizure onset zones (SOZ) compared to other areas. However, the detection of these events and their relation to epileptogenesis is still challenging because there is no formal or global definition of HFO. Visual annotations of HFOs are regarded as the gold standard, but they are extremely tedious, inevitably subjective and require a large amount of concentration. Some previously published automatic detectors show promising performance, but they usually require the optimization of several parameters and supervised validation. In order to simplify this procedure and implement automatic detection of HFOs broadly in scalp EEG data, we modified an automatic HFO detection algorithm initially designed for intracranial EEG data and adapted it to scalp EEG data. The algorithm is based on the iterative estimation of the amplitude distribution of the EEG background activity. It requires the optimization of two parameters related to the number of detected events and determines different optimized threshold values for each 1-minute window in each channel using the iterative procedure. Since scalp EEG can be easily influenced by artifacts like muscle movements, sharp waves and fast transients, it is not trivial to remove such artifacts and achieve good performance. Therefore, several post-processing methods were applied to remove spurious detections and reduce the number of false detections. This procedure was completed with the aid of a new interface which provided good visualization of visually marked HFOs and automatically detected events.
After applying the algorithm to an EEG dataset, we examined the distribution of events detected in all channels and found that the detector had much better performance in subjects where HFOs were mainly concentrated in less than five channels and stood out obviously from background activities. Overall, our detector achieved a true positive rate (sensitivity) of 86.3% and false detection rate (FDR) of 11.3%. If better sensitivity was desired, allowing more false detections, the detector provided a TPR of 91.2% and FDR of 31.7%. After doing a leave-one-out cross-validation within all patients, a TPR of 80.4% and FDR of 32.7% were obtained. We consider this algorithm a powerful tool in localizing high frequency epileptic activities in scalp EEG data due to its advantage in high sensitivity, low false detection rate and the implementation of several artifact rejection methods.