Contextual MEG and EEG Source Estimates Using Spatiotemporal LSTM Networks.

John G Samuelsson,Matti S Hämäläinen,Christoph Dinh,Sheraz Khan,Alexander Hunold

doi:10.3389/fnins.2021.552666

Abstract

Most magneto- and electroencephalography (M/EEG) based source estimation techniques derive their estimates sample wise, independently across time. However, neuronal assemblies are intricately interconnected, constraining the temporal evolution of neural activity that is detected by MEG and EEG; the observed neural currents must thus be highly context dependent. Here, we use a network of Long Short-Term Memory (LSTM) cells where the input is a sequence of past source estimates and the output is a prediction of the following estimate. This prediction is then used to correct the estimate. In this study, we applied this technique on noise-normalized minimum norm estimates (MNE). Because the correction is found by using past activity (context), we call this implementation Contextual MNE (CMNE), although this technique can be used in conjunction with any source estimation method. We test CMNE on simulated epileptiform activity and recorded auditory steady state response (ASSR) data, showing that the CMNE estimates exhibit a higher degree of spatial fidelity than the unfiltered estimates in the tested cases.

Highlights

Magneto- and electroencephalography (M/EEG) have excellent sub-millisecond temporal resolution but limited spatial resolution
The Long Short-Term Memory (LSTM) prediction is not able to capture the temporal dynamics of the neural activation patterns adequately; the waveforms get distorted as seen in the second column of Figure 5
This ability of Contextual Minimum-Norm Estimate (CMNE) to extract the signal from noisy data could be the reason for the superior performance on the auditory steady state response (ASSR) data, where the Kalman and mixed-norm estimate (MxNE) approaches performed substantially worse than CMNE in spite of the steady local activity in the auditory cortex that occurs during ASSR where one might think the Kalman and MxNE approaches would do better

Summary

Introduction

Magneto- and electroencephalography (M/EEG) have excellent sub-millisecond temporal resolution but limited spatial resolution. The most commonly used M/EEG distributed source estimation methods, e.g., MNE, dSPM, and sLORETA, are linear and source estimates are derived time-sample by time-sample, without considering the temporal sequence (Hamalainen and Ilmoniemi, 1994; Dale et al, 2000; Pascual-Marqui, 2002). In other words, these methods fit their source estimates directly to the sensor data without assuming any relationship between the neuronal current distributions across time. The ill-posedness of the inverse problem along with the low SNR in M/EEG recordings cause the limited spatial resolution of the MEG and EEG technologies (Samuelsson et al, 2020).

Methods

Results

Discussion

Conclusion