Electrocorticographic Recordings

Electrocorticographic (ECOG) signals were recorded from 65 patients participating in a feasibility clinical trial of the responsive stimulation System. This was a multicenter trial conducted between 2004 and 2007 designed to demonstrate adequate safety and provide preliminary evidence for effectiveness. The trial was approved by the Food and Drug Administration and investigational review boards of each center, and informed consent was obtained from all patients.

The responsive stimulation System provides responsive cortical stimulation via a cranially implanted programmable neurostimulator connected to 1 or 2 recording and stimulating depth and/or subdural cortical strip leads that are surgically placed in the brain according to the seizure focus. The neurostimulator continually senses ECOG activity and is programmed by the physician to detect abnormal ECOG activity and then provide stimulation. Forty-one subjects had leads implanted in the neocortex, 19 had leads implanted in the hippocampus, and 5 had leads implanted in the hippocampus and neocortex. In total, 95 patients had leads located in the neocortex, and 29 had leads in the hippocampus. Two channels of bipolar recordings were available for each lead. Whenever recording sites or the recording montage changed, data from that patient were treated as a new data set, resulting in 146 data sets.

The cranially implanted neurostimulator processes the signals in real time, using detection parameters that were unique to each patient and adjusted over time to detect epileptiform activity. When stimulation was enabled, detection events were followed after a short latency (typically 60–300 microseconds) by electrical stimulation (typically high frequency; >90% of stimulation occurred at frequencies between 100 and 333 Hz, delivered pulses that were 120–200 microseconds wide, and lasted 100–200 milliseconds). ECOG records (sampled at 250 Hz and typically 60–180 seconds in duration) were stored in response to preprogrammed events (such as an individual detection event or multiple consecutive detection events).

Selection of ECOG Data

Some ECOG records contained multiple detection events, each of which would be followed by stimulation if stimulation was enabled. To minimize effects from one stimulation that might affect the analysis of later stimuli, we only analyzed data corresponding to the first stimulation event in each ECOG record. The wavelet analyses described later occurred at discrete time points, which were defined relative to the detection event. In ECOG records containing responsive stimulation, these time points were always measured relative to the detection event immediately preceding the stimulation. This measurement allowed us to compare activity at corresponding time points in ECOG records that contained stimulation (after detection events) and those containing detection events only. During the period immediately following stimulation, signals could be saturated or blanked (ie, no signal available for analysis) or contain transient low-frequency (<1 Hz) artifacts. To ensure that our results were not affected by these sorts of stimulation artifacts, we verified that ECOG signals from every recording channel were nonzero and nonsaturated throughout a window surrounding each time point. This window was large enough to include all time points that might contribute to the analysis via temporal filtering and the wavelet transformation. Any ECOG records that did not meet these criteria were excluded from the analysis. We only analyzed data sets in which a minimum of 10 ECOG records with and without stimulations were available to calculate the phase-locking statistic.

Wavelet Analysis

To study rhythmic activity, we first computed the discrete wavelet transform of activity in each channel at each time point. We based our analysis on a previously described approach. Specifically, for each frequency, f , we first band-pass filtered recordings between f ± 2.5 Hz ( Fig. 1 B) and then convolved the filtered signal with a wavelet of frequency f to obtain an amplitude and a phase given by

Overview of responsive stimulation and data analysis. ( A ) ECOG signals were recorded using chronically implanted intracranial electrodes. Signals were processed in real time to detect epileptiform activity. When epileptiform activity was detected and the stimulation was enabled, detections were followed by high-frequency electrical stimulation, as shown (of note, the sensing amplifiers were blanked during the brief period of stimulation). To determine acute effects of stimulation, we compared signals 1.0, 0.5, and 0.2 seconds before and 1.0, 1.2, and 2.0 seconds after detection events with stimulation with signals before and after detection events in the absence of stimulation. ( B ) To analyze activity at a particular frequency, we band-pass filtered each signal around that frequency, before convolving with a wavelet to obtain the amplitude and phase at that frequency. ( C , D ) To compute phase locking between a pair of signals, for each ECOG record, we computed the phase difference for those 2 signals and converted all the phase differences (corresponding to different ECOG records) to unit vectors in the complex plane. The amplitude of the vector average of all these vectors measured phase locking.

Measuring Phase Locking

Following the approach described by Lachaux and colleagues, we used phases obtained from the discrete wavelet transformation described earlier to calculate a measure of phase locking. For each pair of recordings from each data set, we computed phase differences, converted these to unit vectors in the complex plane (see Fig. 1 C), and used the amplitude of the average of these unit vectors to measure the degree of phase locking or synchrony at frequency f . We refer to this measure as the phase-locking statistic.

Determination of Statistically Significant Phase Locking

We determined whether a particular value of the phase-locking statistic was statistically significant using bootstrapping. Specifically, we computed 1000 random values of the phase-locking statistic. Each of these was the sum of N unit vectors in the complex plane, each of which had a pseudorandom phase. N is the number of phases (or unit vectors) that was used to calculate the original phase-locking statistic. Then, we obtained a P value by determining how often the actual calculated phase-locking statistic exceeded these pseudorandom values of the phase-locking statistic.