Abstracts

Rationale: Different types of changes in the pattern of seizure progression were recently shown to differentiate responders of closed-loop brain stimulation from non-responders (Kokkinos et al., JAMA 2019). To aid the process of quantitatively evaluating and tracking the progress and evolution of seizures, however, we need computational tools that perform automated segmentation and characterization of seizures. We are developing a tool that performs segmentation on intracranial electrophysiological data collected from patients implanted with the RNS System. Starting with seizures that have been marked for ictal onset, this tool uses statistical methods for detecting change-points at which the frequency content of a seizure significantly changes during the ictal time-course. Methods: We performed segmentation based on the short-time Fourier transform (STFT), or spectrogram, of the seizure time-course, for a single channel of interest. The STFT was first preprocessed to remove extraneous noise and retain only significant peaks in the signal. This was achieved by thresholding the STFT to half of its peak value, within a small time-window, and then repeating this process by sweeping this window over the STFT.The STFT was broken down into six discrete, configurable frequency bins (10 Hz each). The frequency content of the ictal signal within these bins was then used to detect change-points in the seizure time-course. At every time instant, we tested for a change as follows: (1) We considered two equal-sized time windows (2 seconds wide) around the time point of interest; (2) Within each of these time windows, we tested for a change in frequency content, for each of the aforementioned frequency bins using a 2-sample Kolmogorov-Smirnov test (KS test); (3) We repeated the KS test at every time point, and constructed a graph of the p-value of this test over time; (4) p-values were combined over frequency bins using Fisher's method; (5) Finally, change points were marked by finding negative peaks in the signal corresponding to the logarithm of the combined p-value; (6) Segments were extracted as the time period between two change points.We demonstrated the use of this segmentation tool on patients implanted with a Responsive NeuroStimulation System (NeuroPace Inc.). Periods of stimulation were identified and ignored during preprocessing for the purposes of seizure segmentation. Results: The tool identified different stages of the seizure, including instances where the seizure pauses and restarts -- “fragmentation,” as documented by Kokkinos et al. (JAMA 2019). The top of the figure shows the 4-channel RNS data at the start of a seizure. The grayed out regions are periods of RNS stimulation that have been identified and disregarded. The bottom figure shows the STFT of channel 1. Vertical dashed lines indicate change-points. The horizontal solid colored lines indicate the scaled log-p-value of the KS test for each frequency bin. The solid white line is a scaled logarithm of the combined p-value for change detection at each time point. The combined p-value from the KS test performs well in demarcating points at which the frequency content of the seizure changes (see figure). Conclusions: This approach successfully identified change points in the electrophysiological content of ongoing seizures. Further evolution of this computational tool is expected to help develop a quantitative understanding of the long-term effects of RNS. Funding: No funding