Participants
Participants were recruited as part of an ongoing, multi-site collaborative project between 2003 and 2016 (Beth Israel Deaconess Medical Center, Brigham and Women’s Hospital, Children’s Hospital and Massachusetts General Hospital in Boston, and New York University Medical Center, New York and Epilepsy Center of National Institute of Clinical Neurosciences, Budapest, Hungary).
Forty-four patients with medically intractable epilepsy undergoing evaluation with subdural grid and/or intracerebral depth electrodes were implanted with the experimental micro-electrodes. Fifteen patients were excluded due to poor data quality, resulting in twenty-nine patients with good quality data (n=29). The choice of patients for intracranial studies, the location of the clinical electrodes, duration of the implantation and, ultimately, the excision of cortex, were determined entirely on clinical grounds by an independent clinical team without regard for experimental considerations. All patients gave written informed consent to the experimental procedure after a thorough explanation of the risks under procedures monitored and approved by the local Institutional Review Boards at the respective centers (USA) or the National Ethical Board (TUKEB, Hungary) in accordance with the Declaration of Helsinki. Determination of clinical etiology was made by the primary clinical team and, where possible, confirmed through pathological analysis of resected tissue.
Recordings
Two types of laminar multielectrode arrays (LME) were employed depending on the clinical situation (details of electrode construction, recordings and analysis have been published previously44. Both arrays consisted of 24 contacts with diameters of 40 μm and intercontact distances of 150 μm on centers arranged in a line. The contacts on each probe were exposed on the side of a polyimide tube with a total diameter of 350 μm.
The first electrode type was a short laminar probe inserted into the gyral crown beneath a subdural grid electrode (n=26). Care was taken to insert the multielectrode perpendicular to the cortical surface in a location strongly suspected to be within the seizure focus and therefore, most probably resected. The second electrode array was inserted into the lumen of depth electrodes implanted in mesial temporal areas (Ad-Tech Medical Instrument Corporation, Racine, Wisconsin) and extended ~3.5 mm from the tip of that electrode (n=3).
Differential recordings were made from each pair of successive contacts to establish a potential gradient. Recording apparatus including amplifier and filter characteristics were reported44-47. After wideband (DC-10000Hz) preamplification (gain 10x, CMRR 90db, input impedance 1012 ohms), the signal was split into field potentials (filtered at 0.2-500Hz, gain 1000x, digitized at 2000Hz, 16bit) and multi-unit activity (MUA: filtered at 300-5000Hz, gain 1000x, digitized at 20000Hz, 12bit), and stored continuously co-recorded with the clinical electrodes using time-locked triggers.
Analyses were performed using Neuroscan Edit 4.3 software (Compumedics, El Paso, TX) and with custom designed MATLAB (MathWorks, Natick, MA), LabVIEW (National Instruments Corp., Austin, TX) and C/C++ codes, as well as publicly available software suites (EEGLAB)48.
Selection of IIDs
IID detection was performed automatically on daytime LFP traces based on typical morphological characteristics for sharp waves, spikes, and spike-wave discharges49. Artifact-free continuous segments were selected for each patient recorded >2 hours before or after a seizure. Baseline amplitude variance was determined on each channel separately. Activation was considered to be significant when the peak exceeded the 2.5 standard deviation (SD) limit compared to baseline. Each detected event was reviewed visually and those were selected for further analysis that showed biphasic or triphasic morphology with an initial fast phase of maximum 200ms, followed by a prolonged, slower phase that lasted longer than 200ms.
Selection of HFOs
Detection of HFOs was based on the methods described by Staba and Crepon et al. and combined both automated and visual approaches19,22. After careful artifact rejection, putative HFO events were selected automatically by applying a 5SD threshold to the root-mean-square (RMS; 5ms sliding window) transformed bandpass filtered (zero-phase shift 4th order Butterworth digital filter 80-500Hz) data. To avoid the unnecessary detection of the same event recorded on multiple channels, co-occurring (within less than 100 ms) events were grouped on each multielectrode and the highest amplitude event in each group was selected for further inspection. These events were also visually reviewed by 2 epileptologists (DF, ET).
Wave parameters were calculated automatically according to Staba and collegaues22. Duration and number of cycles were determined by the length and number of detected peaks, respectively, of the RMS data exceeding the 3SD threshold. Instantaneous frequency was calculated by dividing the number of cycles by the duration in seconds. Amplitude was expressed as z-score relative to the mean of total artifact-free band-pass filtered data.
The originally selected events were only accepted if they satisfied the following criteria: having at least 3 cycles above the 3SD duration window of RMS data and lasting for more than 6ms. HFO density was given as the number of HFO events in every minute of the artifact-free period (rate/minute).
Since there is a confusion in the literature about the appropriate boundary separating slow and fast ripples, we have addressed this question, by testing multiple band-pass filters (zero-phase shift 4th order Butterworth digital filter) with consecutively increasing lower frequency thresholds (150, 200 and 250 Hz, respectively) to determine the optimal cutoff frequency for the differentiation of slow and fast oscillations.
CSD and MUA calculation
CSD is a powerful method to estimate the laminar distribution and time course of the sources and sinks of membrane currents from LFP recorded at different depths50. CSD was calculated from the second spatial derivative of the LFP signal, as described by Ulbert et al.44 CSD averages were calculated after CSD transformation of the filtered (80-500 Hz) signal. MUA, as a measure of population cellular activity, sheds light on the local cellular responses in relation with the LFP. MUA was obtained by further filtering and rectification (500-5000 48db/oct, zero phase shift) of the signal in a range found to detect action-potentials in previous studies51. A final continuous estimate of MUA is derived by passing the signal through a 50 Hz low pass digital filter (24db/oct zero phase shift)44. MUA significance was calculated from the mean MUA values channel by channel. MUA activity was considered significantly increased if it was higher than mean+5SD and decreased if lower than mean-5SD.
Time-frequency analysis
Two time-frequency methods were applied for high frequency spectral analysis.
1.) Continuous time frequency decomposition was used to detect the ongoing variances of high frequency content of the IID associated field potential gradients using Morlet wavelet transformation (80-500 Hz)52.
2.) Event based frequency analysis for IIDs and HFO peaks (Figure 1). The high frequency content of the marked events, either IIDs or HFOs, was measured event by event based on non-averaged event-related spectral perturbation (ERSP)-maps (using pop_newtimef function of EEGLAB)48, where the event coincided with the time point of peak amplitude determined by the automatic HFO/IID detection. ERSP parameters were set as: baseline: -500 to -200ms in each event, bootstrap significance: 0.001, analysis frequency 80-500Hz. ERSP results were given in decibel relative to the baseline activity. Maximum dB high frequency ERSP time point was selected based on wide range (80-500Hz) frequency average of the ERSP map.
Frequency measurement time window was set to -100 to +100 ms relative to this maximum. Average frequency content in this window was calculated using time averaged ERPS values of the event. All the local peaks were detected in this average frequency curve. Peak frequencies and corresponding decibel values were stored and collected for all the events. The collected decibel values were analyzed as a distribution and plotted on a dB histogram (with bin size = 40). Individual decibel threshold was set based on the upper 5% of this histogram. All the frequency peaks associated with intensity lower than this threshold were rejected. The remaining frequency peaks were collected as the measures of significant frequency content of the marked events.
This analysis was repeated for all the recording channels, and channel versus frequency histograms were calculated based on the detection occasions falling in 10Hz bins on each channel for each type of events in each patient.
Timing of different HFO or IID components
To determine the timing of the HFO, or IID, the steepest slope of the HFO envelope or steepest slope of the IID was used. The HFO envelope was found by calculating the RMS in 5 ms window (11 points) of the band-pass (80-500 Hz) filtered signal. The steepest slope (inflection point) was determined by selecting the local maxima of the first derivative of the HFO envelope or IID in time. If the same event was detected on multiple channels, the onset time of the event recorded in the channel with the highest amplitude was taken into consideration
Recorded cortical layer validation
Due to clinical considerations, histological reconstruction of the electrode penetration in the brain tissue was not always available. Therefore, the sampled cortical depth validation relied on the electrophysiological characteristics of the data derived from previous findings44. Cortical surface was determined by a sudden drop of IID amplitude and an appearance of slow artifacts indicating the surrounding corticospinal fluid.