Abstracts

Rationale: High frequency oscillations (HFOs) have been shown to be correlated both temporally and spatially with seizures. However, the difficulties associated with processing these oscillations make it difficult for them to be used in clinical practice. Due to their low amplitude, short duration, visual identification of ripple (80-250Hz) and fast ripple (250-500Hz) patterns in continuous intracranial EEG (iEEG) is cumbersome. Most work in the field has focused on very limited features of HFOs. We aim to improve detection of these patterns by developing analysis tools based on a semi-supervised clustering method exploring the time-frequency content of HFO. Methods: Human iEEG data was recorded at 2 kHz in five refractory seizure patients over several days. We identified 10-minute segments in sleep, wake and pre-ictal states in each subject and for each day and each seizure. Following high pass filtering the data in an 80-500 Hz range, we computed the RMS of the signal in 100 ms windows and used a median operator to select a robust threshold to capture high frequency events. For each detected event, an epoch of 128 ms before and after the center of the oscillation was extracted and stored as HFO candidates for analysis. Over 230 thousand HFO candidates were identified by the initial threshold, which was set to 3 times of the median variance. In the next step each candidate was investigated in time-frequency plane by using a short time Fourier transform, with a 64 ms sliding window. Several features were extracted from time and time-frequency domain such as line length, peak and median frequency, spectral centroid, and local to global energy ratio. These features were used with a semi supervised X-means clustering. The clustering is executed with a step wise procedure where a k-means clustering is executed at nodes which chosen to be explored by the supervisor. Specifically, a 2-mean clustering is executed at each step and the clusters identified were presented to the expert with their temporal and spatial patterns. Those clusters which were noise or arbitrary spikes were eliminated by the expert and further clustering is executed on remaining clusters to identify HFO patterns. This procedure was executed in each sleep baseline (SB), waking baseline (WB) and pre-ictal segments. Results: In all subjects, we observed that one of the initial two clusters was mostly noise. Both HFOs and spikes were grouped into the same cluster. After the elimination of the noise cluster, the second cluster with HFOs and spikes were further explored. Consistently in all subjects, the second step separated most of the HFOs from spikes and sharp waves. Interestingly, in 4 of 5 subjects the HFO cluster predicted the seizure onset zone. In WB, SB and pre-ictal segments, the average proportions of HFOs detected in seizure channels are 82.94%, 71.50%, and 69.64%, respectively. Conclusions: HFOs are complex dynamic phenomena that are difficult to identify visually in long term iEEG. Our results indicate that semi-supervised methods exploring the time-frequency content of HFOs can be efficiently used to predict the seizure zone.