A Single-Channel and Non-Invasive Wearable Brain-Computer Interface for Industry and Healthcare

The proposal of this work contributes to the development of low-cost, wearable, and portable brain-computer interfaces by exploiting consumer-grade equipment, straightforward signal processing, and extended reality. This technique aims to bring brain-computer interface technology closer to daily life and open new possibilities for many users in both industrial and healthcare applications. The proposal system has also been applied in robot-based rehabilitation for children with attention-deficit/hyperactivity disorder or for autism.

The results were encouraging. To begin, wear the smart glasses and the headband and connect the low-cost electroencephalograph to a PC via a USB cable while the PC is disconnected from the main power supply. At this step, all the electrodes must be disconnected from the electroencephalograph acquisition board to start from a known condition.

At this phase, the EEG stream is processed offline on the PC with a script compatible with the processing implemented in the Android application. Start the script to receive the EEG signals and visualize them. Check the displayed signal.

This must correspond to only the quantization noise of the EEG amplifier. Connect the first electrode and apply the passive electrode to the left ear with a custom clip or use an ear clip electrode. The output signal must remain unchanged at this step because the measuring differential channel is still an open circuit.

Connect an active electrode to the negative terminal of the differential input of the measuring EEG channel and apply it to the frontal region with a headband. After a few seconds, the signal should return to zero. Connect the other active electrode to the positive terminal of the differential input of the measuring EEG channel and apply it to the occipital region with the headband.

A brain signal is now displayed corresponding to the visual activity measured with respect to the frontal brain area and the occipital region. Repeatedly stimulate the user with 10 hertz and 12 hertz flickering icons by starting the flickering icon in the Android application. Press on the touchpad of the smart glasses while also starting the EEG acquisition and visualization script.

Ensure each stimulation in this phase consists of a single icon flickering for 10 seconds. From the ten second signals associated with each stimulation, extract two features by using the fast Fourier transform, the power spectral density at 10 hertz and 12 hertz. Alternatively, consider second harmonics as well.

Use a representation of the acquired signals in the features domain to train a support vector machine classifier. Use a tool in MATLAB or Python to identify the parameters of a hyperplane with an eventual kernel. Based on the input features, the trained model will be capable of classifying future observations of EEG signals.

Disconnect the USB cable from the PC and connect it directly to the smart glasses. Insert the parameters of the trained classifier into the Android application. The system is now ready.

The low-cost electroencephalograph was characterized with respect to linearity and magnitude error. The results are shown here. The flicker of the smart glasses was measured to highlight the eventual deviations from the nominal square wave path.

The characterization of the commercial smart glasses in terms of the amplitude spectrum of the flickering buttons is shown in this figure. Flickering at 10 hertz and at 12 hertz are shown here. This figure represents the signals measured during visual stimulation in the features domain.

The signals associated with the 12 hertz flickering stimuli are presented in blue while the signals associated with the 10 hertz flickering stimuli are presented in red. For each subject, the results associated with a ten second stimulation are compared with those associated with a two second stimulation. The accuracy obtained by considering all the subjects together as well as the mean accuracy among all the subjects are reported here.

The comparison of the classification performance when considering two PSD features versus four PSD features for the SSVEP-related EEG data is shown here. Given the good metrological properties of the consumer-grade equipment, one must pay mostly attention to the mechanical stability of the measuring electrodes because they do not employ conductive gels. This procedure proved functional for SSVEP signals which are relatively robust noise, but one might investigate the usage of similar instrumentation in further paradigms such as motor imagery.

Thanks to the wearability, portability, and easiness of use, this technique is now investigated as a supplementary device for rehabilitation or as a new tool for industrial scenarios.