Deep learning-based EEG source imaging is robust under varying electrode configurations

Objectives

Previous research has underscored the necessity of high-density EEG for accurate and reliable EEG source imaging (ESI) results with conventional ESI methods, limiting their utility in clinical settings with only low-density EEG available. In recent years, deep learning-based ESI methods have exhibited robust performance by directly learning spatiotemporal brain activity patterns from data.

Methods

This study investigates the impact of EEG electrode number on a newly proposed Deep Learning-based Source Imaging Framework (DeepSIF). Through computer simulations and clinical data analysis, we assess ESI performance across various channel configurations (16, 21, 32, 64, and 75 channels) comparing DeepSIF and conventional methods against the simulated ground truth and clinical reference regions.

Results

Our results indicate that DeepSIF consistently delivers accurate source localization and extent estimations across different channel counts and noise levels, surpassing conventional methods. In a cohort of 27 drug-resistant epilepsy patients, the average 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 using 75/16 electrodes.

Conclusions

Our results indicate the robust performance of DeepSIF algorithm for source imaging with low-density EEG.

Significance

Our findings suggest broad applications of the deep-learning based source imaging in clinical settings without the need for high-density EEG devices.