Graph Guided Modulo Recovery of EEG Signals

TL;DR

This paper introduces GraphUnwrapNet, a graph neural network approach for recovering EEG signals from folded observations, leveraging graph structures to improve robustness and accuracy in the challenging modulo recovery problem.

Contribution

The work presents a novel graph-based neural network with a pre-estimation guided feature injection module for improved EEG modulo recovery.

Findings

Outperforms traditional optimization methods in EEG recovery.

Achieves competitive accuracy with current deep learning models.

Demonstrates robustness on the STEW dataset.

Abstract

Electroencephalography (EEG) often shows significant variability among people. This fluctuation disrupts reliable acquisition and may result in distortion or clipping. Modulo sampling is now a promising solution to this problem, by folding signals instead of saturating them. Recovery of the original waveform from folded observations is a highly ill-posed problem. In this work, we propose a method based on a graph neural network, referred to as GraphUnwrapNet, for the modulo recovery of EEG signals. Our core idea is to represent an EEG signal as an organized graph whose channels and temporal connections establish underlying interdependence. One of our key contributions is in introducing a pre-estimation guided feature injection module to provide coarse folding indicators that enhance stability during recovery at wrap boundaries. This design integrates structural information with folding priors into an integrated framework. We performed comprehensive experiments on the Simultaneous Task EEG Workload (STEW) dataset. The results demonstrate consistent enhancements over traditional optimization techniques and competitive accuracy relative to current deep learning models. Our findings emphasize the potential of graph-based methodology for robust modulo EEG recovery.