Editorial: Ictal Source Localization from Intracranial Recordings

Epilepsy is one of the most common neurological disorders in the world. It is characterized by unprovoked seizures, affecting around 70 million people globally (Thijs et al., 2019). About one-third of these patients are drug-resistant and may need to undergo surgical or neuromodulation treatments to control seizures. In these cases, successful surgical intervention relies on the localization of the epileptogenic zone (EZ). Such diagnosis is typically based on examination of physical symptoms or semiology and an extensive presurgical work-up, including multiple noninvasive and invasive techniques. The noninvasive phase usually involves magnetic resonance imaging (MRI), positron emission tomography (PET), video synchronized electroencephalography (video-EEG), etc. This process prepares for a direct surgical treatment or a following invasive diagnostic phase if the noninvasive data is inconclusive (Zijlmans et al., 2019).

As yet, the gold standard to localize the EZ is to identify the seizure onset zone (SOZ) using prolonged intracranial EEG (iEEG) recordings, such as electrocorticogram (ECoG) and/or stereo-EEG (sEEG) implanted to target the hypothesized EZ. In this process, not only the SOZ depicting the origin of ictal activities, but also the irritative zone (IZ) corresponding to the interictal epileptiform discharges (IEDs) are delineated. It has been shown that the IZ is generally not specific to the EZ and are nonunitary in nature (Engel et al., 2009). When iEEG electrodes cover the EZ, intracranial ictal recordings provide an excellent biomarker to delineate the EZ or SOZ, given its high signal-to-noise ratio (SNR), low signal attenuation, and depth spatial coverage. On the other hand, iEEG electrodes have limited spatial coverage and if missing the EZ, the outcome of seizure localization based on examination of iEEG recordings may potentially lead to unsatisfactory surgical resection. This represents a challenge in clinical management of drug-resistant epilepsy, as the contacts of iEEG electrode array cover only a small portion of the three-dimensional space, even though iEEG recording electrodes may be partially covering or missing but close to the actual epileptogenic tissues.

In the work of (Satzer et al., 2022), the authors developed an approach to use electrophysiological source imaging (ESI) techniques to localize the EZ from the ictal sEEG and assess the validity by comparing to the epileptologist-defined SOZ and the surgical outcome. The methodology of ESI is originally developed and commonly used for noninvasive recordings, such as EEG and MEG (magnetoencephalography), to estimate neural electrical activity underlying the noninvasive electromagnetic measurements (He et al., 2018; Michel & He, 2017). Due to the various artifacts caused by movement during convulsion, low SNR, and signal attenuation across the skull, the ictal imaging is nontrivial in the noninvasive modalities. Until recently, new development in ESI techniques combined with advanced signal processing techniques made progress to tackle these issues and demonstrated the efficacy of noninvasive ictal imaging with high fidelity (Nemtsas et al., 2017; Sohrabpour et al., 2020; Yang et al., 2011; Ye et al., 2021). Source localization using invasive electrophysiological recordings represents an interesting complement to identify the EZ in clinical settings.

In this study (Satzer et al., 2022), the authors demonstrated an important application of the ESI methods to inversely reconstruct the ictal source from sEEG recordings, which takes the advantage of three-dimensional spatial covering of the contacts to localize the SOZ. In this work, the authors retrospectively studied 68 seizures from 27 patients. They first identified the dominant frequency at each ictal onset and implemented both rotating dipole fitting and current density estimation methods for source localization. The rotating dipole method was calculated with two signal configurations, i.e., the bandpass-filtered sEEG (bandpass dipole) as the time domain signal and the fast Fourier transform derived frequency bin of interest (FFT dipole) as the frequency domain signal. The current density models implemented two well-established methods: the sLORETA method (standardized low resolution tomographic analysis) and its variant SWARM method (sLORETA-weighted accurate minimum norm). Both models were only computed from the frequency domain signal. It should be noted that the frequency domain signal is a complex number, thus the source localization was operated on the real and imagery components separately. After that, the estimated dipole location and the source distribution were assessed by comparing to the clinician defined information, i.e., the SOZ location indicated by sEEG contacts, and the treatment volume (TV) associated with seizure-freedom and non-seizure-freedom outcomes.

Overall, all ESI methods showed good efficacy comparing to the SOZ or TV. More specifically, the median distance to the SOZ was 7 mm for the bandpass dipoles and 6 mm for the FFT dipoles. The current density model achieved up to 86% accuracy in predicting the SOZ. When comparing to the TV, or the presumed EZ in those seizure-free patients, the bandpass dipoles was significantly closer to the TV in the seizure-free versus non-seizure-free patients. Further analysis suggested that the bandpass dipoles could predict the treatment outcome with sensitivity of 91%, specificity of 63%, and accuracy of 74% under an optimal setting. However, such efficacy was not found in the FFT dipole and the current density models.

Interestingly, in a subset of the cohort, the authors also demonstrated that such source localization results were strongly dependent on the near-field signals. In other words, when the signals close to the presumed pathological sites were excluded and only far-field sEEG recordings were used, the localization results to the SOZ degraded for about 10 mm in about 40% patients, though the fitted dipole location could still track either the SOZ or the TV with a reasonable accuracy under group-level statistics. The overall analysis suggests that the near-field recordings of the pathological tissue is necessary for the sEEG-based approach, in performing good source localization and effective prediction of the treatment outcome for the frequency analysis based source localization. This also indicates that a good pre-sEEG hypothesis is the premises for such approach to achieve favorable performance. Further simulation analysis recommended a nearest distance of 15–30 cm from the source to the sEEG recording contacts. Previously, it was suggested, in a simulation study, that combing both sEEG and EEG improves accuracy of source localization, where scalp EEG can balance the information contents that may be missed when sEEG electrodes are far from the EZ (Hosseini et al., 2018).

In another aspect, there are multiple limitations in this study as well. The current density models were only constructed with the frequency domain signals, while it might be beneficial to extract temporal basis functions dominant to the ictal activities prior to source reconstruction using the time domain signals (Sohrabpour et al., 2020; Yang et al., 2011; Ye et al., 2021). The far-field analysis provides interesting evidence; however, the cohort size was limited and thus may be insufficient to draw strong conclusions. Also, as the authors have pointed out, the observations about the necessity of the near filed recordings may be limited by the source localization methods adopted. Given that the approach was mainly based on narrow band extracted temporal and spectral signals, recent developments in spatiotemporal ESI approaches, which incorporate more information from the time and frequency domain, may provide potential tools for exploration (Sohrabpour et al., 2020). On the other hand, invasive recordings are intrinsically recording both local and distant signals from the brain, reflecting a combination of volume conduction and propagation of the electrical sources (Gonzalez-Martinez, 2016). This makes the spatial interpretation challenging, considering the specific listening zone of the recording contacts (McCarty et al., 2022). In this regard, it remains unknown whether the recorded sEEG from near and far field were dominated by the conducted signals, since the physical principles of source localization models lie in the assumption of volume conduction (Michel & He, 2017). In the case of propagational signals, another recent study may provide a contemporary approach by mapping the delayed or propagated ictal signal forming travelling waves to localize the SOZ (Diamond et al., 2021).

In general, this study tackles an important direction for source localization from invasive ictal recordings, given the current multifarious developments in source localization and exploration using various epileptiform biomarkers. It opens a window to the current clinical applications for accurate localization of the EZ when noninvasive modalities are inconclusive. It would be worthy to bring up attention to source localization using biomarkers beyond spikes, such as seizure and high-frequency oscillations (Cai et al., 2021), from noninvasive and invasive modalities.