The Extent of an Epileptogenic Zone: Application of Signal Processing Methods

Introduction and definitions

The delineation of an effective surgical strategy depends on an accurate delineation of the epileptogenic zone (EZ); this is not a trivial matter. Whereas the EZ has been defined theoretically as “the area of cortex that is indispensable for the generation of epileptic seizures,” recognized post hoc by the attainment of a seizure-free outcome and distinct from the seizure-onset zone (SOZ), Bancaud and Talairach developed an earlier operational definition. Here, “primary organization” is foremost, this referring to the spatio-temporal dynamic structure of the ictal discharge. Stimulation-induced seizures (see chapter elsewhere in this book) are crucial to informing this organizational construct, but the question raised in this chapter is whether data from spontaneous seizures alone are sufficient for defining the EZ extent.

According to SEEG methodology, additional zones are understood. The most relevant for consideration here is the “primary” irritative zone (IZ), the cortical region(s) that display interictal abnormalities epileptiform (spikes or pathological high frequency oscillations [HFOs]) which colocalize—by definition—with the EZ; remote irritative regions are referred to as “secondary.” ^, An important line of investigation is whether the primary IZ, and thus the EZ, can be defined independent of ictal data. Identification of such a biomarker, especially if measurable noninvasively, would considerably simplify the presurgical workup. To date, the most robust biomarker, HFOs, has not proven adequate for this task.

Various signal processing methods may be used to supplement visual analysis. These may be broadly categorized into time-frequency methods, which aim to depict the frequency content of a signal as a function of time, and connectivity methods, which aim to describe various interareal relationships. An important caveat with connectivity measures is that most techniques assume linearity and stationarity of the time series, conditions that do not hold for ictal data, though nonlinear methods may be appropriately employed. Connectivity measures will not be discussed in this chapter.

The dilemma of fast activity

Low-voltage fast activity (LFVA) has long been considered a hallmark of the EZ, and surgical outcomes are improved when it is demonstrated on intracranial recordings. As discussed below, seizure-onset patterns (SOPs) that contain LFVA are prognostically favorable. A vexing dilemma is whether all (early) fast activity is equally localizing of the EZ, and if not, how to discriminate these. For example, the temporal-plus epilepsies were identified as a distinct electroclinical entity because of the surgical failure in patients for whom regions of “very fast propagation” were not included in the resection. Yet, the presence of fast activity in the insula per se is not a determinant of surgical prognosis , that is, not all patients with (early) fast activity in the insula have a temporo-perisylvian epilepsy. Moreover, fast activity at onset is frequently encountered in homotopic cortical regions contralateral to the EZ, especially in mesial frontal cortices. Discriminating “primary” from “propagated” fast activity remains a challenge, and methods that rely solely on fast activity risk inaccurate localization of the EZ.

The dilemma of latency

Latencies are often assumed to denote epileptogenicity, and the boundaries of the SOZ are defined by latency. Yet, the temporal criterion (e.g., one second ) that demarcates this boundary is inherently arbitrary, and most studies broaching this question have utilized a subdural approach. Moreover, while there are clearly measurable latencies in mesial temporal seizures, neocortical seizures more typically demonstrate quasi-simultaneous onset across multiple regions. Indeed, “network organization” is the more common observation. Further, when one considers the microdomain (where seizures truly start), activity may be late to appear in the macrodomain. This important scale consideration aside, temporal sampling (the time when activity is recorded) is inherently dependent on spatial sampling; one must also always consider the “missing electrode.” To posit a relevance in sub-second relative latencies across channels is to posit a perfect spatial sampling. Therefore, how seizures start—SOPs/dynamics—may be more important than when seizures start (relative latencies). Primary organization may be more important than “early spread.”

Time–frequency techniques

Methods exist for measuring and quantifying the frequency content of a time series. The most often employed, and utilized in the methods described below, are short-time Fourier transform, continuous wavelet transform (CWT), and Hilbert transform. The mathematical basis for these will not be elaborated, but each serves to capture the amount of frequency content (magnitude/power) as a function of time. Thus, one may quantify the frequency changes (spectral dynamics) that occur with time; these are displayed as a spectrogram (or scalogram for the CWT).

Analysis is typically done on bipolar channels to eliminate extraneous common mode signals; however, this demands an important consideration. As bipolar recordings are inherently susceptible to in-phase cancellation, amplitude, and thus magnitude/power cannot be used to directly compare spectral content across channels. Furthermore, high frequencies are of inherently lower amplitude (the so-called 1/f spectrum ). Thus, to adequately quantify high frequencies and compare spectral content across channels, some form of normalization is required, most often against a “baseline” of interictal data.

Quantification of fast activity and latency: Epileptogenicity index

The Epileptogenicity Index (EI) was introduced to quantify the emergence of fast activity in temporal lobe epilepsies. First, an energy ratio of the power within the high- and low-frequency bands is defined as a function of time: ER = (Eβ + Eγ)/(Eθ + Eα). This serves to capture the transition to LVFA. Second, a change-point algorithm (CUSUM) is used to indicate when this ratio deviates. Essentially, a penalty term is introduced to negatively offset the ER, which is then cumulatively summed. A detection threshold is set to identify the inflection point of this cumulative sum. Both detection terms are adjustable, and the process is thus semi-automated. The detection time is used to scale the ER. Finally, the values are normalized such that the highest channel is 1. Using this methodology ( Fig. 7.1 ), Bartolomei et al. quantified the relative epileptogenicity of structures within the temporal lobe in a cohort of patients. Structures with an EI > 0.3 were considered highly epileptogenic. In a subsequent study, temporal lobe networks were identified by a clustering algorithm (k-means) according to relative epileptogenicity by structure. At the group level, there was an association between the number of structures with high epileptogenicity (EI > 0.3) and seizure-free surgical outcome.

The EI methodology has since been applied to multiple epilepsy etiologies and anatomic types. More recently, EI has been used as a surrogate for the EZ extent, finding that interictal HFO rate was inadequate (low sensitivity) to define the EZ, and that surgical outcomes were improved when the EZ (EI > 0.4) was completely removed. It is unclear what portion of these complete-resection patients had a temporal lobe resection, or what portion had a “focal organization,” itself associated with improved surgical outcomes.

The EI was defined and optimized for temporal lobe epilepsies, but its performance in extratemporal lobe epilepsy has been questioned. The problem arises in seizures with a quasi-simultaneous onset across multiple cortical regions. In this situation, the latency term effectively drops out of the calculation, and EI thus depends solely on the ER. A recent study demonstrated that further scaling the ER by the Euclidean distance between contacts, essentially penalizing propagation, outperformed standard EI in term of conforming to the resection bed in seizure-free patients, especially in neocortical epilepsies. This modified method also implicated areas outside of the resection in patients with poor surgical outcome (EZ incompletely resected). Obviously, the Euclidean distance between contacts carries little biological relevance, which the authors acknowledge.

Mapping fast activity: Epileptogenicity mapping

A related method relying on fast activity was introduced by David et al. This utilized an approach, statistical parametric mapping (SPM), usually applied to functional imaging data, and was accordingly named epileptogenicity mapping (EM). Power in the fast frequency band (60–100 Hz) was quantified by CWT, normalized against a baseline, and using the spatial positions of electrodes, linearly interpolated to produce a statistical map after thresholding (SPM). Given the inherently spatial approach, comparison of the EM to resection volumes is feasible. As mentioned above, EM was applied to a cohort of temporal lobe patients, mapping the early (<10 seconds) involvement of the insula; this carried no prognostic significance.

A more heterogeneous, mostly extratemporal, cohort was recently published, comparing the EM with the postoperative resection mask. While the “FASTectomy” ratio was higher in seizure-free patients at the group level, this was achieved with only a partial resection of fast activity (mean 28%). Discrimination was not achieved at the patient level: some non-seizure-free patients showed higher resection ratios than seizure-free patients. That is, EM alone cannot be used to direct surgical planning. As with EI, this method appears to poorly discriminate between primary and propagated fast activity. Both methods also assume that only fast activity is relevant for EZ localization.

Consideration of slow frequency: Quantified frequency analysis index

Ictal infraslow activity (baseline shift/DC shift) has recently gained interest in SEEG analysis. ^, Gnatkovsky et al. developed a method that combined three quantitative parameters: 1. fast activity power (>80 Hz) 2. slow polarization shift (SPS), the integral of the smoothed time series after application of a moving average filter, and 3. flattening, intended to capture the transition to low voltage that occurs with LFVA. These three parameters were combined into a single index score. Unlike the above, this method was applied to monopolar (i.e., referential) data, presumably because infraslow activity (SPS) is better quantified in a referential montage. The analytic window was set as the entire seizure. A threshold index score was found to separate EZ from non-EZ contacts in a retrospective cohort. Applied prospectively, QFAI agreed with visual evaluation in 12 of 14 patients; the postoperative outcome was reported to be favorable though the follow up was short (2–14 months).

The above three quantitative approaches were compared in a small study of four patients. In the three seizure-free patients, QFAI most closely matched the surgical resection. Both EI and EM tended to indicate regions that were not within the EZ. The fourth case was not operated upon; here, the methods diverged greatly.

The QFAI method depends on the coincident emergence of ictogenic features: transition to fast activity, background suppression, and superimposed on baseline shifts. The analysis is applied to the entire seizure, not merely the period of onset. Infraslow activity shows late localizing value, with a second prominent phase at seizure termination, ^, which may account, in part, for why this method outperforms EI/EM. Furthermore, the relevance of SOPs/dynamics becomes evident, not merely the tracking of fast activity.

Seizure onset patterns/dynamics: Transition matters

Recent studies have described a taxonomy of SOPs, ^, with common observations but some variability in terminology. SOPs that contain LVFA, whether as the initial change or following a train of “preictal” spiking (or otherwise), convey a favorable surgical prognosis. Despite the appellation “preictal” these spikes are part of the seizure. Facilitated by CWT scalograms, Wu et al. described these dynamics, finding that patterns with greater transitional complexity better localized the EZ (seizure-free resection). Specifically, the occurrence of preictal spiking with superimposed fast activity (gamma/HFOs) followed by suppression was most localizing (all patterns contained LVFA by inclusion). Transitional patterns did not depend on the presence of lesion ^, ; hence, the authors concluded that perilesional cortices with these dynamics were part of the primary organization of the seizure.

Fingerprint of the EZ

The dynamics of the preictal–ictal transition were also considered by Grinenko et al. A pattern of features unique to the EZ (seizure-free resection) was identified from baseline-normalized CWT scalograms. These were: 1. preictal spiking, 2. band-limited (vs. broadband) fast activity, and 3. coexisting suppression of lower frequencies at the time of LFVA. The fast activity bands were typically multiple and importantly, were not harmonically related. These features were found to occur coincident in channels within the EZ; none of these features alone was localizing ( Fig. 7.2 ). Moreover, neither the maximal frequency nor its timing (latency) was discriminating of the EZ at the patient level. Various feature characteristics (29 in total) were extracted, and a support vector machine (SVM)—a commonly used supervised machine learning algorithm—was trained to classify channels as EZ or not. This identified the EZ in 15 of 17 patients, with a false-positive rate (FPR; channels predicted as EZ outside of the resection) of 0.7%.