Abstracts

Rationale: Non-invasive EEG source imaging (ESI) is an important technique (He et al., 2018) to localize and image epileptic activities with high spatial and temporal resolution, aiding in presurgical planning. While high-density EEG (HD-EEG) generally offers benefits for both clinical assessment and scientific research by providing data from more channels compared to low-density EEG (LD-EEG), in many cases the latter is adopted in clinical settings due to practical constraints. This study explores the impact of low-density EEG on non-invasive epilepsy source imaging. By investigating the effects of EEG electrode numbers on ESI performances, this study addresses the question whether robust ESI solutions can be obtained from low-density EEG by means of deep learning algorithms.

Methods: This study investigates how different electrode configurations impact the performance of DeepSIF (Sun et al., 2022), a deep learning-based source imaging framework using biophysically constrained deep neural network, with channel count 16, 21, 32, 64, and 75. We evaluated DeepSIF's performance through computer simulations and in 27 drug-resistant epilepsy patients, comparing results with surgical resection outcomes. Additionally, we compared DeepSIF with established ESI methods including sLORETA and LCMV beamformer. We also investigated the performance of DeepSIF across various challenging conditions, including different signal-to-noise ratios (SNRs) and source locations (depth). The performance is assessed using metrics that consider both the distance and overlap between the estimations and ground truths.

Results: The results show DeepSIF's robustness across varying electrode numbers and SNR levels, outperforming sLORETA and LCMV methods in source imaging. DeepSIF maintains low localization errors across all electrode configurations ranging from 16 to 75 electrodes. While sLORETA and LCMV exhibit increased errors with fewer electrodes, DeepSIF's performance is consistently accurate. Among a group of 27 drug-resistant epilepsy patients, the mean spatial dispersions for DeepSIF, sLORETA, and LCMV are 7.9/9.0 mm, 21.9/28.1 mm, and 20.0/28.9 mm, respectively, when employing 75/16 electrodes. Fig. 1 illustrates an example of source imaging of an averaged interictal spike using DeepSIF, sLORETA and LCMV, in a patient who was seizure-free 1 year post-surgery.

Conclusions: Our results suggest that DeepSIF shows robustness in localizing epileptiform activity from low-density EEG, offering implications for EEG source imaging with a limited number of electrodes, such as routine EEGs in clinical settings.

Funding: NIH R01NS127849 and R01NS096761, and T32 EB029365.