Linear SVM Algorithm Optimization for an EEG-Based Brain-Computer Interface Used by High Functioning Autism Spectrum Disorder Participants

Abstract

Machine-learning algorithms can be used for data classification on EEG-based Brain-computer interfaces (BCIs). Here, we used an algorithm based on linear support vector machine (SVM) to identify the presence of the P300 component in datasets from 15 young adult participants with autism spectrum disorder that were provided for the IFMBE Scientific Challenge 2019. We optimized the parameters and inputs for a linear SVM model throughout the ten attempts of the challenge and compared them in terms of accuracy. The highest score (mean accuracy) of 82% was achieved by a procedure that was customized per session per participant. When using a similar procedure for classification model generation and configuration of parameters for all sessions and participants, the highest score achieved was 77%. The results showed that adding data from targets from different calibration sessions from the same participants to the training dataset resulted in a significant increase in accuracy. In all attempts, the mean accuracy was above 70%, which is considered the minimum classification level for the controllability of a BCI. These are promising results for future use of BCIs as a tool for attention training in ASD participants.