Optimizing Prediction Model for a Noninvasive Brain-Computer Interface Platform Using Channel Selection, Classification, and Regression.

A brain-computer interface (BCI) platform can be utilized by a user to control an external device without making any overt movements. An EEG-based computer cursor control task is commonly used as a testbed for BCI applications. While traditional computer cursor control schemes are based on sensorimotor rhythm, a new scheme has recently been developed using imagined body kinematics (IBK) to achieve natural cursor movement in a shorter time of training. This article attempts to explore optimal decoding algorithms for an IBK paradigm using EEG signals with application to neural cursor control. The study is based on an offline analysis of 32 healthy subjects' training data. Various machine learning techniques were implemented to predict the kinematics of the computer cursor using EEG signals during the training tasks. Our results showed that a linear regression least squares model yielded the highest goodness-of-fit scores in the cursor kinematics model (70% in horizontal prediction and 40% in vertical prediction using a Theil-Sen regressor). Additionally, the contribution of each EEG channel on the predictability of cursor kinematics was examined for horizontal and vertical directions, separately. A directional classifier was also proposed to classify horizontal versus vertical cursor kinematics using EEG signals. By incorporating features extracted from specific frequency bands, we achieved 80% classification accuracy in differentiating horizontal and vertical cursor movements. The findings of the current study could facilitate a pathway to designing an optimized online neural cursor control.