Dynamic Mapping of Ictal Source Propagation From Stereo-EEG

Commentary

The central challenge in intracranial EEG is to reconstruct seizure onsets from sparse sampling of electrodes placed in regions that were presupposed to be involved. This is typically done on a visual basis. Clinicians look not for a single isolated event, but for a pattern that evolves over space and time, then try to mentally infer the onset based on this evolution. A common ictal pattern begins as subtle, low-voltage fast activity that evolves into a rhythmic discharge and then repetitive spiking, spreading from 1 electrode contact to adjacent and then more distant ones. This visual identification of a dynamic, spreading pattern is the benchmark for localizing a seizure. The manual reconstruction of complex patterns of 3D propagation can be difficult and subjective, especially when the electrodes are not precisely at the onset, or when the onset appears broad or involves multiple regions.

Rather than visual inspection of raw EEG, this challenge can be addressed by quantitative methods for source localization. The difficulty in identifying onset visually is mirrored by a well-known “ill-posed” nature of source localization. ¹ Many (technically infinite) configurations of the underlying source can produce the same electrical potentials on EEG. For interictal discharges, this limitation is often addressed by reducing the solution space to the simplest possible model, a single dipole source. More complex methods for volumetric sources apply other mathematical constraints, or regularization, to reduce the solution space, which in some ways simulates what clinicians do in visual analysis. EEG and MEG techniques such as minimum norm estimation, ² standardized low-resolution electromagnetic tomography (sLORETA), ³ and iteratively reweighted edge sparsity minimization (IRES) ⁴ use spatial regularization to the physically “smoothest” or most constrained source. These methods provide static snapshots of activity and fail to directly incorporate the temporal evolution that is a cornerstone of visual identification and localization.

Thio et al ⁵ introduce a novel algorithm, temporally dependent iterative expansion (TEDIE), that is specifically designed to reconstruct neural activity patterns that propagate and expand over time. Rather than solving for independent source reconstructions at each point in time, such as the methods above, TEDIE uses the reconstructed source at each time point as a starting guess for the next, and minimizes a cost function that simultaneously considers how well the EEG data is estimated at each point along with a measure of “smoothness” over time. This temporal regularization should ensure solutions expand and move in a continuous, physically plausible manner, effectively generating a 3D “movie” of the seizure's expansion path.

The authors demonstrate that this method yields more temporally stable and accurate localizations than static methods. They validated the model on simulated ictal data with known ground truth; TEDIE outperformed sLORETA and IRES. They then analyzed intracranial data from 46 patients, most of whom underwent resection or ablation. Removal of the TEDIE-identified propagation zone was a statistically significant predictor of seizure freedom, suggesting that understanding spatiotemporal propagation is crucial in ensuring good outcomes, matching clinical experience. ⁶ Among the 23 patients who were not seizure-free postoperatively, in 12 patients, the TEDIE-identified onset zone had not been removed. A major strength is the paper's careful sensitivity analysis. Results were robust to changes in referencing, electrode density, removal of individual electrodes, onset time selection, and cutoff threshold. Multiple simultaneous sources were also allowed.

An important caveat is that the predictive performance was driven almost entirely by extratemporal cases. In temporal lobe epilepsy patients, resection of the TEDIE source was not a significant predictor of surgical outcome (n = 21, p = 0.48), while in extratemporal lobe epilepsy it was highly significant (n = 25, p = 0.01). This discrepancy highlights limitations that require further study before such techniques are ready for widespread clinical implementation.

There are some potential sources of bias in the model. The initial source estimate depends on a seed based on the intracranial electrode with the highest initial gamma power. The selection seemed robust as using different “runner-up” starting points produced nearly identical maps; however, it is not clear how this would fare in cases in which many electrodes appear involved at once. Additionally, seed selection was based on 30–100 Hz gamma activity, but not all seizures begin this way. The model assumed that seizures propagate isotropically, and anisotropic head models based on diffusion imaging did not appear superior in cases in which such imaging was available. It could be informative, however, to look at this in temporal lobe epilepsy cases specifically, as the lack of connectivity information might conceivably contribute to the relatively poor predictive performance in these cases, given the temporal lobe's long white matter connections to both nearby and distant structures.

Notably, TEDIE is not the only source localization method to incorporate temporal information. Most closely related to TEDIE are spatiotemporal regularization methods, which more directly extend spatially constrained source localization models. One example is fast adaptive spatio-temporal IRES, which optimizes a cost function that incorporates both spatial and temporal smoothness. ⁷ However, this method was designed for fast events (interictal spikes) and thus identifies a single global solution over the time window of the spike, rather than the sequential tracking approach used to follow propagating sources over time. Kalman filtering, an engineering technique that has been applied to EEG, ⁸ uses a dynamical system approach to predict a source's state at the next time point, which is then updated by the actual measurement. The method is robust to noise and missing data and is efficient enough to be used in real-time applications. Dynamic causal modeling is a network-based method that identifies causal interactions between different predefined brain regions, but it relies on a predefined hypothesis of the network structure, rather than being purely data-driven, such as localization algorithms. ⁹

TEDIE and similar dynamic approaches mark a critical evolution in the quantitative analysis of ictal patterns in intracranial EEG. There has been a paradigm shift in source localization from static snapshots to models that fundamentally incorporate ictal propagation over time. Although further refinement and validation will be needed, for example, by incorporating anatomical connectivity data, tools incorporating these ideas should soon be incorporated into the clinical review of intracranial EEG in the same way that quantitative EEG methods have become a standard part of critical care EEG. Such tools have the potential to offer a clearer, more reliable understanding of individual patients’ seizure dynamics than is evident from visual inspection of raw signals alone.