A Review of Language BCIs:Comparing Input Methods and Machine Learning Models

Language BCIs are promising technologies that offer alternative ways of communication for people with disabilities. The most common language BCIs can be classified by their input methods as invasive or non-invasive and by their underlying machine learning methods as supervised learning or deep learning. The supervised methods examined include LDA and SVM. The deep learning methods, including CNNs and RNNs, are explained and discussed. Many other studies have reviewed those methods broadly or focused on comparing particular mechanisms; however, this review details and compares research into language BCIs specifically. This review aims to explain the mechanisms and nuances behind frequently studied language BCI methods classified based on those two criteria, discuss each of their strengths and limitations, and anticipate future trends. This review identifies the current gap in BCI research as the tradeoff between fidelity and invasiveness when it comes to input methods and the lack of data-efficient, transparent machine-learning algorithms.