Spike ripples localize the epileptogenic zone best: an international intracranial study

Abstract

We evaluated whether spike ripples, the combination of epileptiform spikes and ripples, provide a reliable and improved biomarker for the epileptogenic zone compared with other leading interictal biomarkers in a multicentre, international study.

We first validated an automated spike ripple detector on intracranial EEG recordings. We then applied this detector to subjects from four centres who subsequently underwent surgical resection with known 1-year outcomes. We evaluated the spike ripple rate in subjects cured after resection [International League Against Epilepsy Class 1 outcome (ILAE 1)] and those with persistent seizures (ILAE 2–6) across sites and recording types. We also evaluated available interictal biomarkers: spike, spike-gamma, wideband high frequency oscillation (HFO, 80–500 Hz), ripple (80–250 Hz) and fast ripple (250–500 Hz) rates using previously validated automated detectors. The proportion of resected events was computed and compared across subject outcomes and biomarkers.

Overall, 109 subjects were included. Most spike ripples were removed in subjects with ILAE 1 outcome (P < 0.001), and this was qualitatively observed across all sites and for depth and subdural electrodes (P < 0.001 and P < 0.001, respectively). Among ILAE 1 subjects, the mean spike ripple rate was higher in the resected volume (0.66/min) than in the non-removed tissue (0.08/min, P < 0.001). A higher proportion of spike ripples were removed in subjects with ILAE 1 outcomes compared with ILAE 2–6 outcomes (P = 0.06). Among ILAE 1 subjects, the proportion of spike ripples removed was higher than the proportion of spikes (P < 0.001), spike-gamma (P < 0.001), wideband HFOs (P < 0.001), ripples (P = 0.009) and fast ripples (P = 0.009) removed. At the individual level, more subjects with ILAE 1 outcomes had the majority of spike ripples removed (79%, 38/48) than spikes (69%, P = 0.12), spike-gamma (69%, P = 0.12), wideband HFOs (63%, P = 0.03), ripples (45%, P = 0.01) or fast ripples (36%, P < 0.001) removed.

Thus, in this large, multicentre cohort, when surgical resection was successful, the majority of spike ripples were removed. Furthermore, automatically detected spike ripples localize the epileptogenic tissue better than spikes, spike-gamma, wideband HFOs, ripples and fast ripples.

In a large collaborative study, Shi et al. show that spike ripples are more effective than other leading biomarkers at revealing brain areas that generate seizures. An automated spike ripple detector performed well in surgical patients across centres, providing a new tool to guide surgical care in patients with severe epilepsy.

Introduction

Epilepsy is the most common neurological disorder, affecting 50 million individuals globally and accounting for 1% of the global disease burden.¹ One-third of people with epilepsy have persistent seizures despite pharmacological interventions. For subjects with drug-refractory epilepsy, neurosurgical interventions, including resection and neuromodulation, are the most effective treatments. The success of both surgical interventions relies on accurately identifying the epileptogenic zone (EZ)—the brain tissue responsible for generating seizures.

Interictal spikes and high frequency oscillations (HFOs) are two widely used biomarkers to identify the EZ.^1,2 These biomarkers have been extensively studied but have shown poor specificity in clinical practice. Spikes—high amplitude, sharp, voltage fluctuations—are specific for pathology, but not for the EZ.^3,4 Ripples—brief, low-amplitude bursts between 80–250 Hz—are spatially focal⁵ but not specific for pathology since similar physiological events are present during normal cognitive processes.^6-8 Higher frequency (>250 Hz) fast ripples correlate with the EZ in subjects with epilepsy but are less frequently observed.^9-11 Although epileptiform spikes and ripples are understood to represent separate neurophysiological events, nearly half of the ripples co-occur with spikes.^12-16 By combining the pathological specificity of spikes and spatial specificity of ripples, the co-occurrence of ripples and spikes (spike ripples) may yield a more specific biomarker for the EZ compared with either biomarker alone.^7,17-24 For example, a subset of spikes—those with preceding gamma activity (spike-gamma)—was found to localize the EZ with good sensitivity.²⁵ Given the combined advantages and disadvantages of spikes and ripples, spike ripples thus may provide a better biomarker for the EZ than other leading automatically detected interictal biomarkers.

To evaluate whether spike ripples provide a reliable and improved biomarker for the EZ compared with other leading biomarkers, we tested three main hypotheses. First, that the majority of automatically detected spike ripples would be removed in subjects who were seizure free after epilepsy surgery. Second, that the subjects who were seizure free after resection would have a higher proportion of spike ripples removed compared with those who were not seizure free. Third, that the subjects who were seizure free after resection would have a higher proportion of spike ripples removed compared with the proportion of other leading interictal biomarkers of the EZ—spikes, spike-gamma, wideband HFOs, fast ripples and ripples—at both the group and individual levels. To test these hypotheses, we analysed an international multi-institutional dataset of intracranial recordings in subjects with known surgical outcome and >1 year of follow-up using previously validated approaches to automatically detect each interictal biomarker.

Materials and methods

Subject selection

In total, 109 subjects with drug-refractory epilepsy who underwent intracranial phase II evaluation, surgical resection and at least 1-year follow-up were included from four tertiary epilepsy centres (n = 39 site A, n = 12 site B, n = 44 site C, n = 14 site D). We note that data and detection approaches from three prior studies evaluating ripples, fast ripples and HFOs as predictors of the EZ were included in this study,^26-28 in addition to any subsequently available datasets from subjects that met our inclusion criteria from these centres through 12/19/2021. The determination of the seizure onset zone and the surgical plans were made independent from the original studies that collected these data^26-28 and prior to information from this study being available.

All enrolled subjects were scored using the International League Against Epilepsy (ILAE)²⁹ scale to assess their surgical outcomes. For the majority of subjects (n = 104), the channels in the resected volume (RV) were determined by co-registering post-resection MRI with pre-resection MRI to find the overlap between resected channels and the resection cavity. In five subjects, a post-resection MRI was not available, and the resected channels were determined at the local site (prior to this study) by using a combination of the pre-resection MRI, post-implantation CT and the operative’s plan and report. In subjects who achieved successful seizure control after surgery (ILAE 1 subjects), the RV provides the best approximation of the EZ, although we note that even in these carefully selected subsets, the RV usually contains more brain tissue than just the EZ (overlying tissue to reach deep lesions, for example).

This study was approved by the Institutional Review Boards of Montreal Neurological Institute, University of Freiburg, Mayo Clinic and University of Michigan.

EEG recording/acquisition

All sites recorded with their local equipment. For two sites, data included recordings from 1–3 a.m. on post-implantation Days 4–7. For the remaining two sites, data included an epoch containing slow-wave sleep on post-implantation Day 2. Data from sites A and site B were recorded using a Harmonie acquisition system (Stellate), with a sampling rate of 2000 Hz. Data from site C were recorded using a Neuralynx acquisition system (Bozeman) with a sampling rate of 32 kHz and a 9 kHz antialiasing filter. Data from site D were recorded using a Natus Quantum acquisition system with a sampling rate of 4096 Hz and a 1.2 kHz anti-aliasing filter. Data from the Neuralynx and Natus Quantum acquisition systems were downsampled to 2500 Hz and 2048 Hz, respectively, using the decimate function in MATLAB (MathWorks, Natick, MA, USA). Conventional macroelectrodes were used at all sites. Stereotactically placed depth electrodes were used at all centres. Subdural grid and strip electrodes were used at sites A, C and D. The site A and site B recordings included periods selected for slow wave sleep. The site C and site D recordings included approximately 2 hours of interictal data recorded from 1–3 a.m.

We characterized intracranial electrode type as ‘depth’ if depth electrodes were used or ‘subdural’ if subdural grid or strip electrodes were used. A subset of subjects with both depth electrodes and grid or strip electrodes were grouped based on which recording type accounted for most electrode contacts.

EEG preprocessing

All data were referenced to bipolar differences between adjacent channels for a more localized representation of brain activity. An experienced electroencephalographer visually identified and manually removed channels with no signal or high noise. Periods of low-quality recordings caused by intermittent or continuous artefacts exceeding 30 s were also rejected, as well as all periods containing electrographic or electroclinical seizures.

To identify high amplitude artifacts, we detected all voltage peaks in the absolute value of the signal; identified those peaks that exceeded 10 times the standard deviation (STD) of the mean peak amplitude for each channel; and removed ±1 s of data centred on each detected peak. The choice of 10 times the STD to identify artifacts was based on visual inspection of the high-amplitude artifacts within the raw data. To remove recording breaks, characterized by a discontinuous change in voltage, we computed the second differences between adjacent data-points in voltage for each 1 s non-overlapping epoch (an approximation of the second derivative). The difference between the minimum and maximum values of the second differences were computed, and epochs in which this difference exceeded 200 μV/s² were removed. The unfiltered, artefact-free data were then evaluated by the automated spike ripple detector.

Automated spike detection

Spikes were detected on all data from all sites using Persyst 14 software (Persyst Development Corporation, San Diego, CA, USA) with the built-in Reveal algorithm,³⁰ which enables channel-specific spike detection. Visual analysis of several detections confirmed accurate detections.

Automated spike-gamma detection

Spike-gamma (elevated gamma activity in the 0.5-s period preceding a spike) were detected on all data from all sites using the detector as described by Thomas et al.^25,31 The EEG data were processed using MATLAB software (MathWorks, Natick, MA, USA).

Automated HFO detection

Wideband HFOs were detected on all data from all sites using the fully automated ‘quality HFO’ (qHFO) algorithm,³² which detects wideband HFOs (80–500 Hz). This method has been validated on intracranial data to have similar yields as human reviewers. The EEG data were processed using MATLAB software (MathWorks, Natick, MA, USA) and the General Data Flow Package.³²

Automated ripple and fast ripple detection

For each subject from site A or B, the rates for interictal ripples and fast ripples were calculated using the exact detections reported by Jacobs et al.²⁷ Only this subset of data was used for the comparison of performance of ripples and fast ripples with spike ripples.

Automated spike ripple detection

Our group has previously developed two separate spike ripple detection strategies for use in scalp EEG: a feature-based algorithm applied to time series data,^18,33 and a convolutional neural network (CNN) applied to spectrogram images.³⁴ To apply these tools to intracranial data, we first retrained the CNN detector on hand-marked events. We then created an intracranial detector that combines both the feature-based and CNN strategies. We validated the resulting detector on the hand-marked events using a leave-one-out cross-validation procedure. The details of these steps are outlined below.

To train and validate the intracranial spike ripple detector, we compiled a diverse training dataset using a subset of data from 18 subjects across the four sites with ILAE 1 surgical outcomes. For each subject, one expert (W.S. or K.W. after consensus training) manually inspected 10 min of data from five EEG channels within the EZ and marked all spike ripple events (n = 1700, across all 18 subjects). We then applied the feature-based algorithm to five channels outside of the EZ for each subject to define 950 false positive detections across all subjects. Finally, we randomly selected 1600, 1 s samples from the five channels outside of the EZ across all subjects as true negative detections. Spectrogram images of 1 s epochs (either containing a true positive, false positive or true negative detection) were generated to train the CNN detector.³⁴ Note that the negative samples used for training the CNN detector in our study comprised both false positive detections and true negative detections, but these two categories were not distinguished during the CNN training process. As reported by Nadalin et al.,³⁴ we applied transfer learning to train the CNN (i.e. we retrained layers of an existing neural network architecture using the spectrogram image training data).³⁵ The CNN returns the probability of a 1 s epoch containing a spike ripple. To perform leave-one-out cross-validation, we trained the CNN on the marked datasets from 17 subjects and then evaluated the performance of a detector using both the feature-based and CNN algorithms on the left-out subject at different CNN probability thresholds. We repeated this process for all subjects in the training dataset and then aggregated the results to evaluate the overall performance of the intracranial detector. The threshold that yielded the best precision was selected to classify the data not used in training.

We subsequently applied the validated intracranial spike ripple detector to the remaining 91 subjects. For the 18 subjects that contributed data used in training, we applied a detector trained on the other 17 subjects to all data across all channels in the left-out subject. Thus, no subject data contributed to both training and detection in this analysis.

Statistical analysis

To compare clinical characteristics among subjects with ILAE 1 versus ILAE 2–6 outcomes, we performed two-tailed t-tests [degrees of freedom (DF) = 107] or Pearson chi-squared tests (DF = 1–10) between distributions or proportions, respectively.

To investigate the effect of recording length on estimates of spike ripple rate, we performed a sensitivity analysis. To do so, we investigated data from two subjects from each site with the longest duration of data available. For each subject, we selected the channel with the highest spike ripple rate and randomly (uniformly) sampled 50 subintervals of data of increasing duration (6–60 min)³⁶ from the full recording. For each subject, we then computed the spike ripple rate of each subsample and the spike ripple rate STD across subjects and subsamples for each recording duration. We reported the duration beyond which the mean STD of the spike ripple rate remained below 0.5 spike ripples/min.

To test the hypothesis that the majority of detected spike ripples were removed in subjects seizure free after surgical resection, we utilized the previously reported metric, the event rate ratio.²⁷ This ratio compares the sum of events detected in resected channels against those in unresected channels using the following equation:

$$\begin{array}{rl}& \mathrm{Event}\mathrm{Rate}\mathrm{Ratio}(\mathrm{R}{\mathrm{R}}_{\mathrm{Event}})\\ & ={\displaystyle \frac{{\displaystyle \sum _{\mathrm{Resected}\mathrm{channels}}}\mathrm{Rat}{\mathrm{e}}_{\mathrm{Event}}-{\displaystyle \sum _{\mathrm{Non}\text{-}\mathrm{resected}\mathrm{channels}}}\mathrm{Rat}{\mathrm{e}}_{\mathrm{Event}}}{{\displaystyle \sum _{\mathrm{Resected}\mathrm{channels},\mathrm{Non}\text{-}\mathrm{resected}\mathrm{channels}}}\mathrm{Rat}{\mathrm{e}}_{\mathrm{Event}}}}\end{array}$$ (1)

where the numerator is the difference between the summed event rates from removed channels and unremoved channels, the denominator is the summation of event rates from all channels, and all channels were observed for the same amount of time. The event rate ratio ranges from −1 to 1. Values >0 in subjects with ILAE 1 outcome indicate that most events were removed and thus co-localized with the EZ, whereas values <0 indicate that the majority of events were not co-localized with the EZ (Fig. 1). We tested if subjects who were seizure free after resection (ILAE 1 outcomes) had a higher proportion of spike ripples resected than non-resected using a right-tailed Wilcoxon rank-sum test (DF = 47), and repeated this test in the subsets of ILAE 1 subjects from the same site (DF = 18,2,15,9) and with the same intracranial electrode type (DF = 21,25). To test for a relationship between the proportion of channels within the EZ of ILAE 1 subjects and the event rate ratio for spike ripples, we estimated a linear regression model of RR_{Spike Ripple} with predictor proportion of channels resected (DF = 45). The effect size (d-values) was computed using Cliff’s delta.

Figure 1.

Illustrations of scenarios with different event rate ratios. Subdural (array of blue circles) and depth (array of rectangles) electrodes monitor brain voltage. Channels with high event rates are indicated in red. The resected region (green shaded area) overlaps with the true epileptogenic zone (yellow shaded area) in subjects with good surgical outcomes. (A) Channels with high event rates are within the resected region and the event rate ratio is 1. (B) Half of the channels with high event rates are within the resected region and the event rate ratio is 0. (C) None of the channels with high event rates are within the resected region and the event rate ratio is −1. (D) In some cases, the epileptogenic zone may not be fully sampled by the intracranial investigation, resulting in a high event rate ratio but poor surgical outcome. Brain schematics are modified from https://pngimg.com/image/86636, under CC BY-NC 4.0 license.

To test the hypothesis that subjects with surgical cure after resection had a higher proportion of spike ripples resected compared with subjects with continued seizures after resection, we evaluated whether the mean of the distribution of event rate ratios for subjects with ILAE 1 outcomes was greater than the mean of the distribution of event rate ratios for subjects with ILAE 2–6 outcomes, using a right-tailed Wilcoxon rank-sum test (DF = 47).

To test the hypothesis that a higher proportion of spike ripples were resected in subjects seizure free after surgery compared with other proposed biomarkers—spikes, spike-gamma, wideband HFOs, ripples and fast ripples—at the group level, we compared the event rate ratios of each biomarker (RR_{Spike ripple}, RR_Spike, RR_{Spike gamma}, RR_HFO, RR_Ripple and RR_{Fast ripple}) in subjects with ILAE 1 outcome by applying pairwise right-tailed exact binomial tests. We used this approach to account for cases with zero detections. To complement these group-level analyses, we also evaluated in subjects with ILAE 1 outcomes the proportion of channels with high event rates resected for each biomarker type. High event rates were defined as the 90th, 95th and 99th quantiles of the event rate distribution for each biomarker.

To test this hypothesis at the individual level, we estimated surgical outcomes in individual subjects using each biomarker. Subjects with ILAE 1 outcomes were considered correctly predicted by the biomarker if the event rate ratio was positive (RR_Event > 0), indicating that the majority of events were removed. We tested if the success rate of estimated surgical outcomes was higher for spike ripples than other biomarkers (i.e. spike, spike-gamma, wideband HFO, ripple or fast ripple) using right-tailed two-proportion z-tests. Cliff’s d-values were also calculated where applicable to quantify the effect size.

Results

The intracranial automated spike ripple detector has high precision

To develop a detector with high precision appropriate for large volumes of multichannel intracranial data, we detected candidate events that satisfied both a feature-based detector that directly analyses time series data and a CNN detector that analyses spectrogram images.^33,34 Using a leave-one-out cross-validation procedure, we determined the probability threshold (0.65) of the CNN detector that achieved the highest precision against detections hand-marked by experts. While the intracranial spike ripple detector identified less than half of hand-marked events (sensitivity, 42%), it had high precision (78.5%) and a low false positive rate (6.7%), resulting in balanced performance (F1-score = 0.68). The detector required approximately 6.8 min to analyse 10 min from 10 channels of data sampled at >2 kHz (1.8 min for feature-based analysis; 5 min for the CNN analysis) on a single processor with 3.8 GHz CPU (64 GB RAM). The time cost includes input/output operations and data preprocessing steps necessary for this detector. Example intracranial spike ripples detected in the resected volumes of ILAE 1 subjects are shown in Fig. 2.

Figure 2.

Example spike ripple detections. (A) Example detected spike ripple, showing the unfiltered voltage recording (top), the 100–300 Hz bandpass-filtered voltage recording (middle) and the spectrogram (bottom). The spectrogram displays the power in decibels and warmer colors represent higher powers. Red shading in each sub-plot indicates the time interval of the detected spike ripple event. (B) Additional examples of unfiltered voltage recordings of spike ripple detections from diverse brain regions including mesial temporal, neocortical temporal, frontal, parietal and occipital lobes. All detections share the same feature of an epileptiform spike co-occurring with a ripple.

Data characteristics

Subjects with drug-refractory epilepsy and known surgical outcomes after 1 year of follow-up were included (n = 109, 59 females; median age 32 years, range 8–65 years). Among these, 48 subjects had ILAE 1 outcomes. Subject clinical characteristics are provided in Table 1. We found no evidence of a difference in age at surgery, epilepsy duration, sex, surgical site, electrode type, or aetiology between those with or without surgical cure (P ≥ 0.13, all tests). Available EEG data from one night was provided from each centre for each subject. The median duration of intracranial EEG data analysed was 120 min/subject (range 10–121 min). Subject recordings included an average of 58 channels (range 11–147). Sensitivity analysis revealed that for durations of data exceeding 4.7 min, the mean estimate of the standard deviation of the spike ripple rate varied by less than 0.5 spike ripple/min (Fig. 3). Thus, consistent estimates of spike ripple rate can be obtained from approximately 5 min of data. We note that the data used in this study had a duration of at least 10 min.

Table 1.

Summary of patient cohort demographics

	Good surgical outcome (seizure free; Engel IA or ILAE 1)	Poor surgical outcome (continued seizure; Engel IB-IV or ILAE 2–6)	P-value
Total number of subjects	48	61
Age at surgery, years, mean (±STD)	31 (±14.6)	34.1 (±12.6)	0.24a
Epilepsy duration
Mean (±STD), years	19.7 (±11.7)	17.8 (±10.9)	0.52a
Unknown	22	29
Sex			0.44b
Male	24	26
Female	24	35
Surgical site			0.20b
Temporal	25	38
Extra-temporal	22	19
Unknown	1	4
Specific surgical sitec			0.18b
Frontal	19	15
Parietal	5	2
Temporal	25	38
Mesial	2	5
Neocortical	1	1
Both	5	6
Not specified	17	26
Occipital	1	0
Insula	0	1
Unknown	1	4
Electrode typec			0.58b
Depth	37	50
Subdural	26	29
Number of channels, mean (±STD)	62.9 (±30.4)	54.6 (±28.5)	0.15a
Number of channels resected, mean (±STD)	24 (±14.1)	19.7 (±15)	0.13a
Proportion of channels resected, mean (±STD) (%)	39.3 (±15.5)	36.8 (±19.9)	0.49a
Aetiology			0.89b
Lesional	43	55
Non-lesional	4	4
Unknown	1	2
Specific aetiologyc			0.38b
Gliotic lesion/gliosis	11	24
FCD	19	11
MTS	5	11
Non-lesional	4	4
Encephalocele	3	3
Tumourd	4	2
MCDe	1	4
Unknown	1	2
Cyst/cavitation	1	1
Encephalomalacia	1	1
TSC	1	1

Patients grouped by the surgical outcome. FCD = focal cortical dysplasia; ILAE = International League Against Epilepsy; MCD = malformations of cortical development; MTS = mesial temporal sclerosis; STD = standard deviation; TSC = tuberous sclerosis complex.

^a t-test.

^bPearson chi-squared test.

^cSome patients included in multiple categories.

^dTumour includes dysembryoplastic neuroepithelial tumour, ependymoma, hypothalamic hamartoma, glioma and oligodendroglioma.

^eMCD includes microdysgenesis, polymicrogyria, periventricular nodular heterotopia and unspecified MCD.

Figure 3.

The spike ripple rate can be accurately estimated from short durations of data. The solid line is the mean standard deviation (STD) of the spike ripple rate estimates (±95% confidence interval) from differing data durations (eight subjects, two per site). The STD of the spike ripple rate estimate remains below 0.5 spike ripples/min for data durations that exceed 5 min. SR = spike ripple.

The majority of spike ripples were removed in subjects with surgical cure at the group level

The majority of spike ripple generating brain tissue was removed in drug-refractory epilepsy subjects who were seizure free after surgical resection (P < 0.001, d = 0.62; Fig. 4). This result was qualitatively consistent across sites (Site A: median RR_{Spike Ripple} = 0.61, P = 0.001, d = 0.56; Site B: median RR_{Spike Ripple} = 0.52, P = 0.35, d = 0.33; Site C: median RR_{Spike Ripple} = 0.71, P < 0.001, d = 0.63; Site D: median RR_{Spike Ripple} = 0.87, P < 0.001, d = 0.8) and in subjects with majority depth electrode recordings (median RR_{Spike Ripple} = 0.61, P < 0.001, d = 0.52) or majority subdural recordings (median RR_{Spike Ripple} = 0.73, P < 0.001, d = 0.69). We note that the relatively high P-value for Site B reflects the lack of statistical power to detect a significant effect due to the small sample size (n = 3), though the event rate ratio indicates that the majority of spike ripples were resected in the majority of ILAE 1 subjects in this small cohort. Among ILAE 1 subjects, the mean spike ripple rate was higher in the RV [0.66/min; 95% confidence interval (CI) (0.60, 0.72)] than in the non-removed tissue [0.08/min; 95% CI (0.07, 0.09); P < 0.001; d = 0.37]. In addition, we found no evidence of a relationship between the proportion of channels within the EZ and RR_{Spike Ripple} (P = 0.22, d = 0.59; Supplementary Fig. 1) among ILAE 1 subjects.

Figure 4.

In subjects with good surgical outcome most spike ripples were removed. Distribution of the spike ripple rate ratio from (A) all subjects with ILAE 1 surgical outcome, (B) from each clinical site and (C) using each intracranial electrode type. Each filled circle indicates one subject. Violin plots⁴⁸ show the density of the spike ripple rate ratio values, the violin width represents the number of subjects at each value, open circles represent median values and darker shading indicates the interquartile range. Data-points that deviate significantly from the majority of each data group are identified as outliers and marked with red crosses. ILAE 1 = International League Against Epilepsy Class 1 outcome.

More spike ripples were removed in subjects with surgical cure than in subjects with persistent seizures. Subjects with curative resection (ILAE 1) had a higher proportion of spike ripple generating brain tissue removed compared with those who were not seizure free (ILAE 2–6; P = 0.06, d = 0.17).

Spike ripples identified the epileptogenic cortex better than other leading interictal biomarkers

Subjects with curative resection (ILAE 1) had a higher proportion of spike ripple-generating brain tissue removed compared with the proportion of removed tissue that generated spikes (P < 0.001, d = 0.16), spike-gamma (P < 0.001, d = 0.40), wideband HFOs (P < 0.001, d = 0.36), ripples (P = 0.009, d = 0.65) or fast ripples (P = 0.009, d = 0.65) (Fig. 5).

Figure 5.

Spike ripples outperform other biomarkers in identifying the epileptogenic zone. Distributions of five different event rate ratios from subjects with ILAE 1 outcome (seizure free after surgery). Violin plots as in Fig. 4. Among these subjects, spike ripples have a significantly higher event rate ratio than spikes, spike-gamma, wideband high frequency oscillations (HFOs), ripples or fast ripples. Significant differences are indicated as: **P < 0.01, ***P < 0.001. ILAE 1 = International League Against Epilepsy Class 1 outcome.

Although the RV in subjects with surgical cure after resection provides the best available estimate of the EZ, the RV may also include channels not necessary for surgical cure. To complement our analysis, we also evaluated whether the majority of channels with high event rates (top 90–99th quantile) were resected. We found that the mean proportion of channels with the highest spike ripple rates resected exceeded the mean proportion of channels with the highest spike, spike-gamma, wideband HFO, ripple or fast ripple rates resected in all quantiles evaluated, with the exception of the 90th quantile, where spike-gamma and spike ripple had similar performance (Table 2). Together, these findings support spike ripples as a more specific biomarker for the EZ than other leading interictal biomarkers.

Table 2.

Mean proportion of resected high-rate channels for each interictal biomarker type in ILAE 1 subjects at different thresholds

Quantile	Spike ripples	Spikes	Spike-gamma	Wideband HFO	Ripples	Fast ripples
90th	67.61%	57.17%	67.86%	61.74%	46.07%	43.86%
95th	74.71%	58.20%	72.48%	71.06%	45.42%	42.40%
99th	77.49%	68.48%	75.00%	75.00%	36.36%	50.00%

HFO = high frequency oscillation.

The majority of spike ripples were removed in most subjects with surgical cure at the individual level

If a biomarker identifies the EZ, we expect its removal to correlate with successful resective surgery. Consistent with this expectation, at the individual subject level, we find the majority of spike ripples were removed (i.e. RR_{Spike Ripple} > 0) in 79% (38 of 48) of subjects with ILAE 1 outcome. The percentage of ILAE 1 subjects with the majority of spike ripples removed was higher than the percentage of subjects with the majority of spikes (69%, P = 0.12), spike-gamma (69%, P = 0.12), wideband HFOs (63%, P = 0.03), ripples (45%, P = 0.01) or fast ripples (36%, P < 0.001) removed. We find consistent results when other thresholds are used to classify epileptogenic tissue (RR_Event > 0.1, 0.2, 0.3, 0.4 or 0.5). In each case, the percentage of ILAE 1 subjects with the majority of spike ripples removed exceeds the percentage of subjects for the other biomarkers (Table 3). These results support the robustness of the finding that spike ripples outperform other leading biomarkers in identifying the EZ at the subject level.

Table 3.

Percentage of ILAE 1 subjects with the majority of events removed for each biomarker

RREvent Threshold	Spike ripples	Spikes	Spike gamma	Wideband HFO	Ripples	Fast ripples
0.0	79%	69%	69%	63%	45%	36%
0.1	79%	60%	65%	60%	27%	18%
0.2	75%	54%	58%	56%	18%	14%
0.3	73%	46%	52%	48%	5%	5%
0.4	69%	44%	42%	46%	0%	5%
0.5	60%	38%	27%	31%	0%	5%

HFO = high frequency oscillation; ILAE 1 = International League Against Epilepsy Class 1 outcome; RR = rate ratio.

Although the majority of spike ripples were removed in 38 of 48 subjects with ILAE 1 outcome, this was not the case in 10 subjects. In four of these subjects, the spike ripple rates were low, with less than 10 spike ripples detected across all channels in each case. In these four subjects, the median spike ripple rate was 0.08/min (mean 0.13/min, range 0–0.27/min). In contrast, the median spike ripple rate in the RV of the other 44 subjects was 1.70/min (mean 4.40/min, range 0.03–28.20). In another four subjects, relatively high rates of spike ripples (>1.1/min) were detected in a hippocampus that was not resected. In the remaining two subjects, we found no unique considerations that may have contributed to biomarker failure.

Overall, these results suggest that spike ripples provide a better estimate of the EZ than other leading interictal biomarkers. However, if spike ripples are detected with low rates across all channels, or at high rates in hippocampi, the biomarker may be less accurate.

Discussion

We evaluated whether spike ripples, the combination of epileptiform spikes and ripples, accurately localized to the epileptogenic zone in an international cohort of subjects who underwent intracranial investigation, surgery and post-surgical follow-up. We found that the majority of spike ripples were removed in subjects who became seizure free after surgical resection and that this finding was consistent across different epilepsy centres and recording electrode types. Additionally, we found that a higher proportion of spike ripples were removed in subjects who were seizure free after resection than in those who were not. Finally, we found that a greater proportion of spike ripples were removed in subjects that were seizure free after resection than the proportion of spikes, spike-gamma, wideband HFOs, ripples or fast ripples. Overall, we conclude that spike ripples better localize epileptogenic tissue than other leading interictal biomarkers and therefore provide a valuable tool for identifying the epileptogenic zone to guide surgical resection.

Previously, we have shown good intra-rater reliability and performance using semi-automated³³ and fully-automated^18,34 spike ripple detection techniques in scalp EEG data. Here we introduce an intracranial detector that extracts both time series features and spectral information to detect spike ripples in intracranial data. We find that this detector has high precision in identifying spike ripples on intracranial EEG recordings. Our findings are consistent with previous work suggesting that spike ripples are less sensitive than spike-gamma but more specific for the EZ (see Table 2 in Thomas et al.²⁵). In prolonged, multichannel intracranial recordings, this detector has sufficient sensitivity to detect nearly half of all events and high precision to limit false detections. This novel automated approach enables objective quantification of spike ripples, providing a reproducible approach to estimate the EZ.

To assess biomarker performance, we measured the event rate ratio, the same metric applied in previous studies to evaluate biomarkers for the EZ; here we included the performance metric, data and detection methods employed in a prospective clinical trial.^27,37,38 We note that large resection volumes can falsely inflate this metric, but this concern is mitigated by the inclusion of multiple centres and surgical approaches and the comparison of different biomarkers using the same metric. Although spike ripples provided a reliable biomarker for the EZ in most subjects studied here, the majority of automatically detected events occurred outside the EZ in eight subjects. Among these subjects, four were found to have low event detection rates among ILAE 1 subjects. This finding raises the possibility that accurate detection of the EZ using this biomarker requires a minimum number of detected events. Although the majority of hippocampal EZ were identified correctly using the spike ripple rate ratio (nine of nine subjects with ILAE 1 outcome and hippocampal resection), the biomarker failed in four other subjects with ILAE 1 outcome who had high spike ripple detection rates in hippocampi that were not resected. These results raise the possibility that the unresected hippocampus may in fact also be epileptogenic if the subject were followed for a longer period of time³⁹ or alternatively that hippocampal sharp wave ripples, which share many features with spike ripples, were misidentified as spike ripples by the detector.^40,41 Until validated methods to distinguish spike ripples from sharp wave ripples are available, limiting spike ripple evaluations to non-hippocampal tissue may be the most prudent. Excluding subjects with low spike ripple rates (<0.3 detections per minute from all channels combined) and those with high spike ripple rates in non-epileptogenic hippocampus, spike ripples accurately localized the EZ in 40 of the remaining 41 subjects, indicating that spike ripples are a highly reliable cortical biomarker when present. We expect spike ripples to be even more informative when integrated with the clinical evaluation. For example, a high spike ripple rate in tissue that is concordant with the clinical hypothesis would strengthen this hypothesis and help demarcate the extent of the required resection, whereas a high spike ripple rate explained by a clear artefact would be ignored. Alternatively, a high spike ripple rate in a region discordant with the clinical hypothesis would indicate that careful re-inspection of this region should be performed and require high confidence in the localizing information provided by other tools to justify ignoring this specific biomarker.

We found a stronger relationship between spike ripples and the EZ compared with other biomarkers (spikes, spike-gamma, wideband HFOs, ripples or fast ripples) and the EZ. For spike detection, we applied the Persyst Reveal spike detector algorithm, which was reported to have a sensitivity of 76% and false positive rate of 0.11/h in scalp recordings.³⁰ The qHFO algorithm was reported to have a specificity of 88.5%.^32,42 The spike-gamma detector was reported to have a sensitivity of 78.9% and a specificity of 69.5%.²⁵ The ripple and fast ripple detector was reported to have a sensitivity of 96.8% and false positive rate of 4.86%.^27,43 We note that the reported sensitivity of the spike, spike-gamma, ripple and fast ripple detector are higher than the spike ripple detector (42%). The specificities of the qHFO and spike-gamma detector are approximately comparable to the spike ripple detector (78.5%). While it is theoretically possible that each of these detectors could be further optimized, this optimization is limited by the inherent features of the events of interest that separate them from the background and the accuracy of the biomarker to reflect disease. Here, different algorithms were used to detect spikes, spike ripples and ripples. However, in each case, we utilized a validated detector, which arguably represents the best available approach for each biomarker. In the case of ripple and fast ripple detection, the original data and detections reported by Jacobs et al.²⁷ were used. In the case of wideband HFOs, the original data and detections were used for data reported by Gliske et al.²⁸ and the same detector was applied to the new data. Thus, while we acknowledge the importance of distinguishing between the accuracy of a biomarker versus a detection algorithm, the two issues are interconnected, since detection accuracy, both using visual analysis or automated tools, is limited by the signal-to-noise relationship of the biomarker to background activity. One of the benefits of spike ripple as a biomarker is the improved ease and therefore accuracy of detecting these combined events, compared with ripples alone.³³ We also note that the results reported here are consistent with our driving hypothesis, that a biomarker that combines the pathological specificity of a spike with the spatial specificity of a ripple will have improved specificity for the EZ compared with either biomarker alone. In addition, we found that spike ripples are an improved biomarker for the EZ compared with other leading biomarkers.

Whether spike ripples reflect a unique event compared with spikes or ripples alone is not yet understood. The observation that spike ripples better localize to the EZ than spikes suggests that these events may be originating rather than propagated discharges. In humans, spike ripples correlate with higher single-unit firing rates compared with spikes.⁴⁴ However, the roles of different cell types (e.g. interneurons) in spike ripple generation remain unclear.^45,46 Understanding the cellular mechanisms that support spike ripples and whether treating these mechanisms impacts seizure recurrence, remains an open challenge.

When interpreting the results reported here, we note the following limitations. First, the epileptogenic zone is a theoretical concept and may not exactly match the resected volume in subjects with good surgical outcomes, as utilized here. An inadequately sampled EZ (as illustrated in Fig. 1D) would result in an uninterpretable event rate ratio using any electrographic biomarker and result in poor surgical outcome. For this reason, the subjects with surgical cure after resection (ILAE 1) provide the best estimate of biomarker performance, since these patients had sufficient sampling for accurate EZ localization. Second, we analysed data from 1–3 a.m. as a proxy for non-rapid eye movement (NREM) sleep for two of four sites, as sleep scoring was not available. Although previous analysis indicated little difference between interictal 1–3 a.m. data and interictal NREM sleep data, definitive NREM recordings would be preferable.^8,47 Additionally, we performed a retrospective analysis of data collected prospectively in a multicentre study evaluating HFOs.²⁷ Future work evaluating the relationship between spike ripples and surgical outcome in a prospective cohort is required to validate these findings. Moreover, the details provided about the failed cases are anecdotal, and caution should be exercised in generalizing these findings. However, we propose that these observations may inform future research to optimize detection approaches and refine surgical decision-making. Finally, as we did not have detailed medication information available for our datasets, we were not able to evaluate the impact of medication changes on spike ripple rate. A future study could evaluate whether attention to medication effects could further improve the specificity of this biomarker.

Overall, we developed a novel automated approach combining feature-based and machine learning algorithms to detect a novel combined biomarker, spike ripples. Using this detector, this study provides evidence that spike ripples are an improved biomarker for detecting epileptogenic tissue compared with spikes, spike-gamma, wideband HFOs, ripples and fast ripples.

Data availability

Derivative data are available from the first author upon reasonable request. Data are available from authors at each site upon reasonable request.

Funding

This work was supported by the National Institutes of Health, National Institute of Neurological Disease and Stroke R01NS119483, R01NS110669.

Competing interests

The authors report no competing interests.

Supplementary material

Supplementary material is available at Brain online.