Recruiting neural field theory for data augmentation in a motor imagery brain-computer interface

Abstract

We introduce a novel approach to training data augmentation in brain-computer interfaces (BCIs) using neural field theory (NFT) applied to EEG data from motor imagery tasks. BCIs often suffer from limited accuracy due to a limited amount of training data. To address this, we leveraged a corticothalamic NFT model to generate artificial EEG time series as supplemental training data. We employed the BCI competition IV '2a' dataset to evaluate this augmentation technique. For each individual, we fitted the model to common spatial patterns of each motor imagery class, jittered the fitted parameters, and generated time series for data augmentation. Our method led to significant accuracy improvements of over 2% in classifying the "total power" feature, but not in the case of the "Higuchi fractal dimension" feature. This suggests that the fit NFT model may more favorably represent one feature than the other. These findings pave the way for further exploration of NFT-based data augmentation, highlighting the benefits of biophysically accurate artificial data.

Keywords: EEG; brain-computer interface (BCI); common spatial pattern (CSP); data augmentation; motor imagery; neural field theory.

FIGURE 1

SES-BCI: The setup includes a DJI-Tello drone, Wearable Sensing DSI-24 EEG headset, personal computer, and an optional virtual reality headset. This system simultaneously executes MI and SSVEP paradigms. EEG signals undergo processing to generate navigation commands for the drone. The drone’s video feed presented to the user includes embedded flickering arrows in the corners, acting as stimuli for SSVEPs. In this setup, the SSVEP paradigm is responsible for forward and backward navigation commands, while the MI paradigm is associated with right, left, up, and down navigation commands. (This figure includes a hospital background image by Pikisuperstar from Freepik).

FIGURE 2

The paradigm sequence utilized in the “2a” dataset from BCI competition IV (Tangermann et al., 2012). We used an EEG segment from t = 2.5 to t = 5 s of each epoch for MI classification.

FIGURE 3

Experimental (inter-epoch average), fitted and simulated power spectra of a right-hand MI CSP source signal between 8 and 30 Hz. Simulated spectra are presented for two distinct parameter values. (A) A reduction in the synaptic decay-time constant α diminishes the total spectral power, especially affecting the high frequencies. (B) A decrease in the cortical damping γ results in a decline in the EEG resonant frequency alpha and its peak shift towards the lower frequencies, along with a slight flattening of the beta peak. (C) A decrease in the corticothalamic propagation delay t ₀ leads to a shift of the EEG resonant frequencies alpha and beta peaks towards higher frequencies. (A–C).

FIGURE 4

On the left panels: model generated time-series of (A) eyes-open resting state, (B) eyes-closed resting state, (C) sleep-stage 2 and (D) sleep-stage 4. On the right panels: corresponding time series from human subjects (Penfield and Jasper, 1954; Nunez, 1995; Robinson et al., 2005).

FIGURE 5

CTM diagram: the neural populations shown are cortical excitatory, e, and inhibitory, i, thalamic reticular nucleus, r, and thalamic relay nuclei, s. The parameter ν _ab quantifies the strength of the connection from population b to population a. Excitatory connections are indicated by pointed arrowheads, while inhibitory connections are denoted by round arrowheads.

FIGURE 6

Power topoplots of CSPs fitted to MI EEG epochs. The fitting process aims to enhance the differentiation between right-hand and left-hand MI conditions.

FIGURE 7

Experimental inter-epoch average EEG power spectrum, analytical power spectrum of a fitted CTM, and a power spectrum calculated from time series simulated by a fitted CTM. Shown are the left-hand MI CSP (A) and the corresponding right-hand MI CSP (B).

FIGURE 8

Workflow of MI data augmentation performance evaluation procedure. The process involves creating a small dataset and assessing accuracy using inverse CV, where one fold is reserved for training and the others for testing. The MI pipeline consists of EEG epoch preprocessing, CSP decomposition, feature extraction, and classification. The NFT-based data augmentation process generates artificial CSP time series using a CTM (Robinson et al., 2002; Robinson et al., 2005; Kerr et al., 2008; Abeysuriya et al., 2015) fitted to the CSP time series from the MI pipeline.

FIGURE 9

Classification accuracy of the TP feature for the full dataset, the small dataset, and the augmented small dataset using different DA strategies. The classification was conducted on validation sets for subjects with high MI proficiency. Noticeably, NFT-based DA approaches show statistically significant improvements in Ac, whereas noise-based DA does not yield significant enhancements.

FIGURE 10

TP (A) and HFD (B) features extracted from CSPs of original and augmented time series. The CSPs were augmented by a factor of 1, employing an NFT model. The left- and right-hand MI trials maintain linear separability.