Exploring Neurophysiological Markers of Brain-Computer Interface Performance in Children
Poster No:
2045
Submission Type:
Abstract Submission
Authors:
(Vella) Shin-Hyung Kim1, Daniel Comaduran Marquez1, Anup Tulhadar2, Arazdeep Minhas1, Joanna Keough1, Eli Kinney-Lang2, Adam Kirton2
Institutions:
1University of Calgary, Calgary, Alberta, 2Alberta Children's Hospital, Calgary, Alberta
First Author:
Co-Author(s):
Introduction:
Brain-computer interfaces (BCIs) are a promising new access technology for children with quadriplegic cerebral palsy who are unable to move or speak (Jadavji et al., 2021). Despite great advances in the field, some users are unable to effectively control BCI systems, leading to the expenditure of valuable time and effort from the user and support personnel that would be better directed towards alternative BCI solutions (Edlinger et al., 2015). Electroencephalographic (EEG)-based metrics may predict adult BCI user's performance but there are no such pediatric studies to date (Blankertz et al., 2010; Ahn et al., 2013; Bamdadian et al., 2014). The aim of this exploratory study was to generate a machine-learning model that best predict BCI performance in children using neurophysiological parameters, and explore features best correlated with success in BCI use.
Methods:
Typically-developing children (n=29, age=10.07±2.19, 58% female) were recruited and attended the BCI4Kids laboratory. 120-180 seconds of resting-state EEG recordings were collected prior to motor-imagery based BCI training and task sessions. The primary outcome of BCI performance was the classification accuracy score of the training sessions. A 60% classification accuracy score from the BCI training was used to classify participants into two groups.
Data from 17 active channels were analyzed. The resting-state EEG data were cropped to 100 seconds and notch-filtered at 60Hz. The following features were calculated from the resting-state EEG recordings for the theta (4-8 Hz), alpha (8-13 Hz), beta (13-30 Hz) bands: 1) The functional network, using the weighted-phase lag index 2) The normalized power spectrum, using the Welch's method 3) The aperiodic components of the power spectrum (offset and exponent), and 4) Graph theory metrics (global efficiency and characteristic path length). These values were averaged between electrodes for the given brain regions: frontal, central, parietal, and occipital. Random forest models (RF) and support vector machine (SVM) models were tested. Five predictive models were generated for each RF and SVM model type, each testing a different subset of EEG features: All features, Power, Connectivity, Aperiodic power, and Graph theory features. Each of these models were generated 50 times, ensuring that the same training and testing sets were used between the models each time. The performance of these models were analyzed using the mean balanced accuracy (BA) and F1 scores. Feature importance was also analyzed.
Data from 17 active channels were analyzed. The resting-state EEG data were cropped to 100 seconds and notch-filtered at 60Hz. The following features were calculated from the resting-state EEG recordings for the theta (4-8 Hz), alpha (8-13 Hz), beta (13-30 Hz) bands: 1) The functional network, using the weighted-phase lag index 2) The normalized power spectrum, using the Welch's method 3) The aperiodic components of the power spectrum (offset and exponent), and 4) Graph theory metrics (global efficiency and characteristic path length). These values were averaged between electrodes for the given brain regions: frontal, central, parietal, and occipital. Random forest models (RF) and support vector machine (SVM) models were tested. Five predictive models were generated for each RF and SVM model type, each testing a different subset of EEG features: All features, Power, Connectivity, Aperiodic power, and Graph theory features. Each of these models were generated 50 times, ensuring that the same training and testing sets were used between the models each time. The performance of these models were analyzed using the mean balanced accuracy (BA) and F1 scores. Feature importance was also analyzed.
Results:
For both SVM and RF models, running a feature subset of all 36 connectivity features showed high mean balanced accuracy scores, but with a high variability in performance (SVM model: BA = 61±19%, F1 = 5.6±1.8; RF model: BA = 61±19%, F1 = 5.6±2.0) (Fig. 1). A feature subset of 15 power features using the SVM model showed the second highest scores, again with high variability (BA = 60±17%, F1 = 5.6±1.6) (Fig. 1). For the power-based RF model, features within the alpha and theta bands in the parietal regions had high feature importance (alpha parietal = 10.4±2.9%; theta parietal = 9.6±2.7%) (Fig. 2A). For the connectivity-based RF model, features within the alpha band in various regions had high feature importance (alpha parietal = 8.3±2.9%; alpha global = 6.9±2.3%) (Fig. 2B).
·Figure 1. Mean balanced accuracy and F1 scores of SVM and RF Models using five different sets of EEG features. Error bars indicate standard deviation.
·Figure 2. Mean feature importance scores of RF Models using (A) normalized power and (B) WPLI connectivity as EEG features. Error bars indicate standard deviation.
Conclusions:
In this exploratory study, we generated predictive models of BCI performance in typically developing children using SVM and RF models for the first time. Models using power and connectivity features had the highest mean performance; however, they are currently variable in their performance. EEG connectivity and power within the alpha band may be important correlates of BCI performance. An increased sample size would better evaluate the models proposed above. Understanding which factors best predict pediatric BCI performance would help users to find optimized BCI strategies, and potentially allow the utilization of these factors to improve a child's performance in BCI tasks.
Modeling and Analysis Methods:
Classification and Predictive Modeling
Connectivity (eg. functional, effective, structural)
EEG/MEG Modeling and Analysis 2
Motor Behavior:
Brain Machine Interface 1
Keywords:
Computational Neuroscience
Data analysis
Electroencephaolography (EEG)
Machine Learning
PEDIATRIC
Other - Brain Computer Interface
1|2Indicates the priority used for review
Provide references using author date format
Bamdadian, A. (2014), 'The Predictive Role of Pre-cue EEG Rhythms on MI-based BCI Classification Performance', Journal of Neuroscience Methods, vol. 235, pp. 138–144.
Blankertz, B. (2010), 'Neurophysiological Predictor of SMR-based BCI Performance', NeuroImage, vol. 51, no. 4, pp. 1303–1309.
Edlinger, G. (2015), 'How Many People Can Use a BCI System?', Clinical Systems Neuroscience, pp. 33–66.
Jadavji, Z. (2021), 'Can Children With Perinatal Stroke Use a Simple Brain Computer Interface?', Stroke, vol. 52, no. 7, pp. 2363–2370.