Steps towards sensitizing EEG feature identification in paediatric brain signals for use in BCIs

Introduction: Brain-computer interfaces (BCIs) designed for paediatrics have yet to be fully realized [1,2]. One hurdle facing paediatric BCIs is the dynamic progression of elctrophysiological states of a child throughout development [3]. Variations in development may be further pronounced in children facing motor or other impairments. One potential solution is to sensitize feature identification to the unique developmental state of a child while accounting for shared patterns across typical development [3]. A proof-of-concept was previously demonstrated on resting-state EEG in paediatric epileptic populations ([4], under review), where EEG signal analysis incorporated tensor (i.e. multi-way) methods in identifying age-specific features. This ongoing proposal aims to replicate our resting-state findings in a paediatric stroke population and to adapt the proposed methods to task-driven EEG, including steady-state visual evoked potentials (SSVEP) and movement-related cortical potentials (MRCPs). Together, these steps could provide critical improvements in signal analysis for EEG BCIs tailored towards paediatrics. Materials, Methods and Results: Two paediatric EEG datasets are used for analysis. First, approximately 30 EEGs from paediatric stroke patients at the Alberta Children's Hospital provides validation of our previous findings in a population ideal for BCI rehabilitation. The goal is to replicate discerning age-specific features in EEGs using tensor analysis methods [4]. Second, publicly available EEG data on healthy participants from the Child Mind Institute (CMI) [5] is explored. Data comprising the high-density 129-channel `Contrast-Change Paradigm' (CCP) of pre-adolescent subjects (approx. 6-11 y.o., n=44) is the focus. The CCP provides a simple motor task, in which users select between left/right images flickering at 20/25 Hz changing in contrast from 50%/50% to 0%/100% (or vice versa) [5]. Due to the CCP design, both SSVEP and MRCP potentials can be isolated, and used as 'simulated-BCI' signals. Tensor analysis exploits the high dimensionality of EEG data to uncover latent relationships in signals, such as age-related features common across subjects. Tensors were constructed with dimensions [Trial]x[Channels]x[Frequency]x[Subject]. Combinations of tensor model decompositions will be explored, including variations on the Parallel Factor (PARAFAC) and Tucker models, to identify dominant underlying factors. Non-negativity and unimodal constraints on different dimensions can improve model stability, interpretation, and help illicit prominent factors across ages. Figure 1 illustrates a PARAFAC decomposition of the CMI SSVEP data, showing the [Channels]x[Frequency]x[Subject] dimensions for left-select (top row,20 Hz) and right-select (mid row, 25 Hz) triggers and their differences (bottom row) at occipital channels. These results are similar in approach to the BCI tensor analysis in [6]. Discussion: The results demonstrate simple tensor analysis, like PARAFAC, is capable of extracting underlying factors associated with signals of interest (yellow), i.e. SSVEP frequencies at 20/25 Hz, while separating out erroneous data (red) and background noise (blue). These separated factors account for properties from each dimensions, including the age-ordered [Subject] dimension, since PARAFAC requires strict 1-to-1 interactions between factors across dimensions [7]. Exploring tensor decompositions across the acquired datasets, along with integrating age-weighted probabilities to the extracted feature profiles, could offer solutions for incorporating age-related features and properties into the signal analysis stage of BCIs. Significance: The details presented here are a work in progress. Adapting the powerful tools of tensor analysis to incorporate developmental information into EEG feature identification lays a framework which could improve BCI signal analysis for children. Replication of this work in both healthy and afflicted paediatric populations will further validate its potential for BCI systems. Acknowledgments: Funding was provided by Edinburgh Neuroscience NeuroResearchers Fund. References: [1]DOI:10.2478/s11536-013-0249-3 [2]DOI:10.1088/1741-2560/13/6/061002 [3]DOI:10.1109/EMBC.2017.8037684 [4]ArXivID:1712.07443 [5]DOI:10.1038/sdata.2017.40 [6]DOI:10.1088/1741-2560/13/2/026005 [7]DOI:10.1016/S0169-7439(02)00089-8