Authors :
Presenting Author: Jesse Rong, – Carnegie Mellon University
Zhengxiang Cai, PhD – Carnegie Mellon University
Boney Joseph, MD – Mayo Clinic
Gregory Worrell, MD,PhD – Mayo Clinics
Bin He, PhD – Carnegie Mellon University
Rationale: High-frequency oscillations (HFOs) are increasingly recognized as promising biomarkers for the delineation of epileptogenic tissue. However, non-invasive imaging of HFOs remains technically challenging due to their brief duration, high frequency content, and low signal-to-noise ratio on scalp EEG. While prior studies have shown the feasibility of deep learning-based source imaging for interictal spikes, there has been limited exploration of applying similar methods to HFOs. This study develops and validates a deep learning framework to accurately localize sources of HFOs from dense array EEG.
Methods:
We developed a transformer-based deep learning framework for non-invasive HFO source imaging using dense array EEG and evaluated its performance in drug-resistant epilepsy patients with known resection outcomes. The model was trained using a large set of biophysically realistic simulations of ripple-band (80–250 Hz) HFOs, generated from neural mass models with focal cortical activation. Pathological HFOs (pHFOs), riding on spikes, were detected from scalp EEG recordings of drug-resistant epilepsy patients using a validated detection pipeline, and source imaging was performed using both the deep learning model and sLORETA for comparison. Results were compared to the surgically resected region, which were quantified using post-op MRIs.
Results:
We analyzed source localization results for 76 pHFO events extracted from 22 patients with both temporal and extratemporal epilepsy. The transformer model demonstrated significantly better performance than sLORETA in terms of precision, localization error, and spatial dispersion (Wilcoxon signed-rank test, p < 0.001), while achieving comparable recall (p > 0.05) (Fig. 1A). Figure 1B illustrates two representative examples of pHFO source imaging results from two patients, who were seizure free 1 year postsurgery. The transformer model consistently localized pHFO sources within or near the resected region, demonstrating strong anatomical concordance. In contrast, sLORETA frequently produced diffuse activations extending beyond the resection zone.
Conclusions: This study demonstrates the feasibility of transformer-based deep learning for HFO source imaging of epileptogenic zone, from dense array EEG. Compared to sLORETA, our model yields more focal, more accurate localizations that better align with resection zones. These findings support the utility of deep learning in advancing non-invasive HFO source imaging and its potential integration into clinical workflows for epilepsy surgery planning.
Funding: NIH NS127849 and NS096761.