Universal Brain-Machine Interfaces Enabled By Flexible Scalp Electronics and Deep Convolutional Neural Networks

Abstract

Electroencephalography (EEG) allows non-invasive monitoring of brain electrical activity at a fine temporal resolution. However, it is difficult to consistently capture high-quality biopotential signals due to natural variations in scalp and skin conditions. Conventional EEG systems suffer from inconsistent signal quality extensive preparation time, and uncomfortable and bulky electrode setups from long, inflexible wires and rigid electrodes. Improvements in analysis and machine learning techniques have allowed researchers to extract more information from EEG with fewer electrodes on the scalp with certain EEG paradigms. However, further hardware improvements would be required before these systems could be comfortably used daily. Here, we introduce skin-like hybrid electronics (SKINTRONICS), a fully portable, wireless, flexible biomonitoring electronics platform incorporating a soft elastomeric membrane for maximum conformability and wearability on the skin. This concept allows for a comfortable and ergonomic wearable BMI concept that can be used long-term without discomfort and is optimized for use in a rehabilitation setting for patients with brainstem injury or other motor impairment. Analytical and computational studies establish the fundamental design criteria of the SKINTRONICS, enabling seamless portable EEG recording with significantly enhanced signal quality over commercial systems. Analysis in the time domain and frequency domain are completed, with deep convolutional neural networks providing appreciable improvements over conventional techniques such as SVM. The real-time time-domain analysis with deep convolutional neural networks allows for highly accurate classification and fine machine control with high information transfer rates from only two channels. With six human subjects this wireless interface achieves a high accuracy (94.54% ± 0.90%) for a corresponding information transfer rate of 122.1 ± 3.53 bits per minute, demonstrated using a wireless wheelchair, motorized vehicle, and keyboard-less presentation using a two-channel EEG. This demonstrates great potential for this platform, which can accommodate a wide range of assistive technologies as well as different EEG paradigms like motor imagery (MI).