Development of a brain-computer interface for patients in the critical care setting

Abstract

Objective: Behaviorally unresponsive patients in intensive care units (ICU) are unable to consistently and effectively communicate their most fundamental physical needs. Brain-Computer Interface (BCI) technology has been established in the clinical context, but faces challenges in the critical care environment. Contrary to cue-based BCIs, which allow activation only during pre-determined periods of time, self-paced BCI systems empower patients to interact with others at any time. The study aims to develop a self-paced BCI for patients in the intensive care unit.

Methods: BCI experiments were conducted in 18 ICU patients and 5 healthy volunteers. The proposed self-paced BCI system analyzes EEG activity from patients while these are asked to control a beeping tone by performing a motor task (i.e., opening and closing a hand). Signal decoding is performed in real time and auditory feedback given via headphones. Performance of the BCI system was judged based on correlation between the optimal and the observed performance.

Results: All 5 healthy volunteers were able to successfully perform the BCI task, compared to chance alone (p<0.001). 5 of 14 (36%) conscious ICU patients were able to perform the BCI task. One of these 5 patients was quadriplegic and controlled the BCI system without any hand movements. None of the 4 unconscious patients were able to perform the BCI task.

Conclusions: More than one third of conscious ICU patients and all healthy volunteers were able to gain control over the self-paced BCI system. The initial 4 unconscious patients were not. Future studies will focus on studying the ability of behaviorally unresponsive patients with cognitive motor dissociation to control the self-paced BCI system.

Fig 1. General scheme of the BCI system used in the experiment.

Fig 2. BCI protocol.

To begin the session, each patient was first presented with the command ‘We are going to start.’ After the command ‘To make it beep faster keep opening and closing your right hand’, the patient was expected to try moving his/her right hand. After the command ‘To make it beep slower stop opening and closing your right hand’, the patient was expected to stop his/her right-hand motion. To ensure patients did not forget the task, each command was repeated every 10 seconds. Duration of each motion/no motion period was randomly chosen from the interval [15, 50] seconds.

Fig 3. Auditory and visual feedback.

(A) Auditory feedback. Each time the motion intention pattern was detected in the EEG signal, the pace of the sound (beets per second, bps) and sound frequency were increased (Δbps = 0.15, ΔkHz = 0.15) until 2 bps and 2 kHz were reached. If the pattern was not detected, both were decreased (Δbps = -0.15, ΔkHz = -0.15) until reaching 0.5 bps and 0.5 kHz. When reaching a pace of 2 bps and a 2 kHz frequency during the motion task the auditory feedback “you have succeeded” was provided to the patient. (B) Visual feedback. Each time the motion intention was detected, the size of the bar was increased by 0.1 until reaching of 1. Otherwise, it was decreased by 0.1, until reaching 0. Motion intention periods were indicated with a green arrow pointing up and a red hourglass was pictured when no motion intention patterns were detected. A smiley face was displayed when the bar reached 1 (full size).

Fig 4. Data formation scheme.

For each time t, to form the explanatory variable x(t), EEG signal from the electrodes was mapped by continuous wavelet transform with 35 frequencies (1, 2, …, 35 Hz). Then, the absolute values of the wavelet coefficients were decimated along the temporal modality to receive 10 points. The response variable y(t) was equal to 1 during the command “Keep opening and closing …” and was equal to 0 during the command “Stop opening and closing …”. The next epoch was taken with a time step of 100 ms.

Fig 5. Signal decoding during calibration and validation sessions of the BCI experiment.

The model was updated every 10 seconds during the calibration session and fixed during the validation session.

Fig 6. Testing results.

(A) Patient 1, (B) Patient 2, (C) Patient 3, (D) Patient 4, (E) Patient 5. Both auditory and visual feedback were provided for the cases of (A) and (B); only auditory feedback was provided for the cases of (C), (D), and (E). For all the cases, value 1 of y-axis corresponds to the task “… keep opening and closing …”, and 0 corresponds to the task “… stop opening and closing …”.