Flexible Silver/Carbon Nanotube-Graphene Oxide-Polydimethylsiloxane Electrode Patch for Electroencephalography Language

Abstract

A flexible silver/carbon nanotube-graphene oxide-polydimethylsiloxane (Ag/CNT-GO-PDMS) patch electrode for recording electroencephalography (EEG) signals and recognizing words is prepared. These patches record EEG signals under the synergistic sensing mechanism of the noncontact capacitance mode of the CNT-GO-PDMS patch and contact current mode of the Ag claws, with low scalp contact resistance of 6.4 kΩ. In the occipital region, the signal-to-noise ratios (SNR) are ≈90 dB for α-waves and 9 dB for visual-evoked signals; the SNR of auditory-evoked EEG signals in the temporal region is ≈10 dB. The EEG cap comprises seven Ag/CNT-GO-PDMS patches to record EEG signals in steady-state visual-evoked potentials (SSVEP) and multiple auditory steady-state response (MASSR). These patches can recognize nine words (“one” to “nine”) in the SSVEP–MASSR paradigm, with a visual accuracy of 90.4% and auditory accuracy of 54.0%. The statistical analysis also shows that the stimulation frequency and scalp channel are significant influencing factors for the accuracy of word recognition. We developed a standardized process of flexible Ag/CNT-GO-PDMS patches and herein propose a new strategy to identify words, which is of great significance for the establishment of the EEG language database and the application of EEG in the field of information transmission.

Introduction

We are now living in the Information Age. Various secrets in business, finance, medicine, and military can be easily captured or decoded using various methods. For example, we need to pass the file password “xxx” to our business partner. The transmission method of saying or writing passwords is at risk of being overheard or watched by others, while the method of sending emails or short messages is at risk of being intercepted by viruses such as Trojan horses. Electronic signals can easily be intercepted or captured. Scalp electroencephalography (EEG) signals are physiological electrical signals expressed on the scalp by the electrochemical reactions of neurotransmitters in the human brain.^[1] These EEG signals can feedback the functional information of the different brain regions.^[2] The medium of EEG-transmitting information is the human brain. Although people as media cannot decode information, they have already transmitted information unconsciously. In addition, the transmission mode of EEG has no intermediate channel, which significantly reduces the possibility of information leakage during transmission. Even if the EEG signal file is captured, the decoded semantics are not transmitted through EEG signal. The semantic recognition method of EEG is newly developed and has not yet formed an EEG language or semantic system. The establishment of EEG language is an important scientific principle applied in brain science and biomedical engineering and has gained attention as a focus area for research in these fields.

At present, the location of EEG recording channels in various functional areas of the brain has been determined.^[3] Taking the 19-lead EEG cap designed according to the electrode position of the 10–20 international system as an example, the frontal lobe related to mental and thinking functions is connected with five electrode channels of Fp1, Fp2, Fz, F3, and F4. The temporal lobe related to auditory function is connected by electrodes in six channels: F7, T3, T5, F8, T4, and T6. The parietal lobe, which controls somatosensory function, is connected externally by six electrodes at the C3, CZ, C4, P3, Pz, and P4 positions. The occipital lobe controls the visual center and memory function, and the connecting channels are O1 and O2. With the increase in the number of electrodes on the EEG cap, the connecting channels of each brain region are also refined, such as the Oz channel in the occipital region, and TP9, FT9, FT10, and TP10 channels in the temporal region.^[4] Therefore, the scalp EEG signals recorded in various brain functional areas using noninvasive scalp EEG electrodes can be decoded to transmit information through the human brain.

To the best of our knowledge, steady-state visual-evoked potential (SSVEP) is a common paradigm for collecting visually evoked EEG signals, while auditory steady-state response (ASSR) can record auditory-evoked feedback EEG signals.^[5] In general, the EEG signal of the SSVEP paradigm has the advantage of high amplitude, and the research results of the SSVEP-brain–computer interface (BCI) systems have been reported in many publications.^[6] For SSVEP, some information can be transmitted by specifying the corresponding relationship between stimulus frequency and words. However, human language and pronunciation are rich, and the number of words is huge. The EEG has an excellent response in the 5–45 Hz range, but the optimal frequency recognition is only 0.2 Hz.^[7] The number of recognizable frequencies of these EEG waves cannot meet the needs of many semantics. In addition, the amplitude of feedback EEG signal is very small in the ASSR paradigm, which is the bottleneck in the field of semantic recognition.^[8]

Materials are the basis for the construction of flexible electrodes and significantly influence electrode performance.^[9] Carbon nanotubes (CNTs) are one-dimensional carbon nanomaterials,^[10] and graphene oxide (GO) is a two-dimensional carbon nanomaterial.^[11] Owing to their advantages of high conductivity, biocompatibility, and electrochemical activity, they are widely used in bioelectric electrodes such as electromyography (EMG),^[12] electrocardiography (ECG),^[13] EEG,^[14] and neural electrodes.^[15] Polydimethylsiloxane (PDMS) is a hydrophobic silicone material that is also widely used to construct various biological interfaces owing to its advantages of good biocompatibility, viscoelasticity, flexibility, and chemical stability.^[12,16] In the early stage, we developed dry- and semi-dry EEG electrodes with graphene, CNT, and PDMS and proved their effectiveness in reducing scalp contact impedance, improving comfort and signal-to-noise ratio (SNR).^[17] Among them, CNT-PDMS-based high-density electrodes were initially applied in the semantic recognition of EEG.^[17a]

In this study, a flexible Ag/CNT-GO-PDMS electrode patch was developed for semantic recognition of brain waves in the SSVEP-multiple ASSR (MASSR) paradigm. The flexible CNT-GO-PDMS film was prepared by spin-coating and then embedded in Ag microclaws. The novelty and significance of this study are as follows. First, the flexible CNT-GO-PDMS film captures a large number of electrons from the scalp in a noncontact capacitive mode, Ag microclaws directly captured electrons from the scalp in contact mode, and the internal CNT-GO conductive network transmits the electrons to the Ag microclaws. Under the synergistic sensing mechanism of such noncontact capacitance and contact current modes, the electrode patches recorded EEG signals with a high signal-to-noise ratio (SNR). Second, the preparation process of flexible Ag/CNT-GO-PDMS electrodes is simple and suitable for standardized batch production, and the cost is sufficiently low to be used as disposable electrode pastes. Third, a new strategy of semantic recognition based on the MASSR paradigm was developed. Fourth, first of all, we realized semantic recognition based on the SSVEP-MASSR paradigm, and nine words (i.e., one, two, three, four, five, six, seven, eight, and nine) were recognized with high accuracy.

Results and Discussions Preparation and Characterization of Ag/GO-CNT-PDMS Electrode Patches

Figure 1 shows a schematic of the preparation and application of the Ag/GO-CNT-PDMS electrode patches. A flexible Ag/GO-CNT-PDMS patch was prepared by spin-coating, cutting, and thermal curing (Figure 1A). The hierarchical structure of the electrode patch is shown in Figure S1, Supporting Information. Our strategy for designing the Ag/GO-CNT-PDMS patches was to harness the synergistic effects arising from the flexibility and stickiness of PDMS, electron capture and transmission of the GO-CNT network, and high electron flow of Ag microelectrodes. In contrast, cooperation between the noncontact capacitive mode and contact current mode was realized. As a result, Ag/CNT-GO-PDMS patches have high skin conformance, comfort, and cleaning convenience, as well as high biocompatibility, low scalp contact impedance, high amplitude, high SNR, and real-time stability. These Ag/GO-CNT-PDMS electrode patches are suitable for developing an EEG semantic recognition system in the SSVEP-MASSR paradigm (Figure 1B).

Figure 1. Schematic of A) the preparation process of flexible Ag/GO-CNT-PDMS electrode patches and B) their application in the acquisition of evoked EEG signals of visual and auditory coupling.

Figure 2A shows the scanning electron microscopy (SEM) images of a CNT-GO-PDMS patch. The patch thickness was approximately 300 μm. The CNT texture was observed in the surface SEM images, whereas the CNT-GO conductive network was observed in the internal SEM images. The GO nano-sheets were curled because they were too thin or were wrapped by the PDMS matrix. In Figure 2B, the SEM image shows an Ag claw in the Ag/CNT-GO-PDMS electrode patch. The Ag claws had a cylindrical shape with a diameter of ≈600 μm (Figure S2, Supporting Information). These Ag claws were mainly composed of Ag atoms and a few C atoms inside the binder. The O, C, and Si atoms distributed around the Ag claw belonged to the CNT-GO-PDMS patch. The Ag claws directly touched the scalp and Cu foil, and the internal CNT-GO conductive network transferred electrons to the Ag claws, thus realizing the dual-mode synergistic effect of the noncontact capacitor and contact current. The Ag/CNT-GO-PDMS patch was confirmed to contain C, O, Si, and Ag elements using energy-dispersive X-ray spectroscopy (EDS) in the SEM system (Figure 2C). In the X-ray photoelectron spectroscopy (XPS) survey spectrum, it can be seen that the CNT-GO-PDMS patch contains C, O, and Si (Figure 2D), and a large number of C═C bonds belonging to the CNTs and GO are present (Figure S3, Supporting Information). Figure 2E shows the Fourier-transform infrared (FTIR) and Raman spectra of the CNT-GO-PDMS patch. Several Si–O–Si and Si–C/Si–CH₃ bonds appeared in the two vibration spectra, as well as the D and G peaks of GO and CNTs in the Raman spectra.^[18] The combination of the Raman and FTIR spectra confirmed the formation of the CNT-GO-PDMS matrix. A 50% CNT-GO-PDMS film was placed in environments with humidity values of 10%–90% for approximately 10 min, and a series of FTIR spectra were obtained (Figure S4, Supporting Information). In the FTIR spectra of the CNT-GO-PDMS films, the broad and intense peaks at 3440 and 1640 cm⁻¹ due to the vibration of –OH suggest that the samples contained a large amount of adsorbed water.^[19] As the relative humidity increased, the absorption band gradually strengthened, indicating that the sensitive film exhibited good hygroscopic properties. This confirms that CNT-GO-PDMS can absorb a large amount of NaCl aqueous solution and sweat. This may be related to the introduction of hydrophilic GO nano-sheets.

Figure 2. A) SEM images of the CNT-GO-PDMS composite film. B) Electron transfer path diagram, SEM image, and corresponding EDS elemental mapping distribution of the Ag/CNT-GO-PDMS composite film. C) EDS spectrum of Ag/CNT-GO-PDMS composite film. D) Survey XPS and E) FTIR and Raman spectra of CNT-GO-PDMS composite.

Selection of Electroencephalography Electrode

Skin-contact impedance, EMG, electrooculogram (EOG), and spontaneous EEG are reference indices for evaluating the interface state between the scalp electrode and the brain.^[20] Figure 3A and Figure S5 and S6, Supporting Information, show the impedance experimental data for optimizing the Ag/CNT-GO-PDMS electrode. The skin-contact impedance collected on the back of the hand (R_hand) first decreased and then increased with an increase in CNT content from 0 to 100%, and the critical point appeared at 50% CNT content. In the range of α-rhythm (8–13 Hz) and bioelectrical (5–1000 Hz) frequencies, the average R_hand values of various CNT-GO-PDMS electrodes are 17.6–24.2 and 10.4–14.9 kΩ, and those of the 50% CNT-GO-PDMS electrode are 17.6 and 10.4 kΩ; the two average R_hand values of the 50% CNT-GO-PDMS electrode with Ag claws are lower, at 13.9 and 10.2 kΩ (Figure 3A). The change rule of the scalp contact resistance (R_scalp) of various electrodes was consistent with that of the R_hand value. The R_scalp range of (0–100%) CNT-GO-PDMS electrodes is 9.66–11.82 kΩ in the hairless area and 11.36–14.3 kΩ in the hair-covered area; the R_scalp of the Ag/CNT-GO-PDMS electrode is 6.4 kΩ in the hairless area and 7.2 kΩ in the hair-covered area (Figure 3B). The R_scalp values of the Ag/CNT-GO-PDMS electrode in the hairless and hair-covered areas met the requirements of EEG signal acquisition.

Figure 3. Impedance and EEG characteristics of Ag/AgCl, Ag/CNT-GO-PDMS, and CNT-GO-PDMS electrodes with different CNT-GO ratios. A) Nyquist plots, frequency–resistance plots, and skin-contact resistances (Rhand) from hand skin. B) Scalp contact resistances (Rscalp) recorded in the forehead and temporal region with hair. C) Physiological electrical signals and D) corresponding frequency spectra recorded in the blinking-eye and teeth-grinding paradigms. E) EEG signals, F) corresponding frequency spectra, and G) time–frequency distributions in a closed/open paradigm. H) Correlations (Rcc) of various CNT-GO-PDMS electrodes with Ag/AgCl electrodes. I) Amplitude and SNR of EEG signals. J) EEG-SNR of the Ag/CNT-GO-PDMS patch in 1 h.

In addition, high-amplitude EOG signals with frequencies of 3–13 Hz were recorded in the blinking-eye paradigm, while high-amplitude EMG signals with frequencies of 3–46 Hz were recorded in the teeth-grinding paradigm (Figure 3C,D and S7 and S8, Supporting Information). Furthermore, the high- and low-amplitude waves generated when closing and opening eyes, respectively,^[21] were recorded using various self-made Ag/CNT-GO-PDMS patches and Ag/AgCl electrodes, with the closed/open-eye paradigm (Figure 3E and S9, Supporting Information). According to the corresponding frequency–amplitude spectrum and time–frequency distribution, the occurrence of 8–13 Hz α-waves in the closed-eye state and spontaneous EEG characteristics without a frequency-doubling response were confirmed (Figure 3G,F and S10, Supporting Information).

The correlation (R_cc) of the EEG signals using a commercial gel Ag/AgCl wet-cap electrode confirmed the reliability of the self-made electrodes. The R_cc values of the CNT-GO-PDMS and Ag/CNT(50%)-GO-PDMS electrodes ranged from 98.63% to 99.95%, with excellent reliability (Figure 3H). The amplitude (5.2 μV) and SNR (86.8 dB) of the spontaneous EEG response of the Ag/CNT(50%)-GO-PDMS electrode were improved by the CNT content and the direct contact effect of the Ag contacts, and these two values were higher than the synchronous data of the commercial electrodes (Figure 3I). The Ag/CNT(50%)-GO-PDMS electrode was connected to the Oz channel of the occipital region of a volunteer, and the spontaneous EEG signals were recorded continuously for 1 h during the process of repeatedly opening and closing the eyes (Figure S11 and S12, Supporting Information). The initial SNR of the EEG signals was 72.1 dB, and the SNR remained at 107.5% in 55 min, with a relative standard deviation (RSD) of 4.7% (Figure 3J).

Both CNT and GO exhibit high conductivity, biocompatibility, and electrochemical activity. In the PDMS elastic matrix, the flexible GO nano-sheet can improve the hydrophilicity, store more sweat, improve the dispersion of CNT, and connect the CNT knots to weave a large long-distance conductive network. In addition, the stress–strain curves of the GO-PDMS, CNT-GO-PDMS, and CNT-PDMS electrodes show that the CNT-GO-PDMS electrodes have the best strain capacity (Figure S13, Supporting Information). This was related to GO improving the dispersion of CNT in the PDMS matrix. The above roles of GO nano-sheet enable flexible CNT-GO-PDMS patches to effectively record EEG signals in a capacitive noncontact mode.

The impedance spectra of the Ag/PDMS, CNT-GO-PDMS, and Ag/CNT(50%)-GO-PDMS electrodes on the skin were recorded (Figure S14, Supporting Information). The volunteer was a 26-year-old woman. The equivalent circuits obtained from the impedance spectra of the three electrodes are R(CR)_S(QR)_S-E(CR)_E, R(Q(CR))_S/E(CR)_E, and R(CR)_S(Q(CR))_S-E(CR)_E, respectively (insets). Generally, in the equivalent circuit of the electrode on the skin, a large resistance appears on the skin, the interface between the skin and the electrode, and the electrode body, which is also the basis for forming subcircuits. According to the equivalent circuit of the Ag/PDMS, CNT-GO-PDMS, and Ag/CNT-GO-PDMS electrodes, a large resistance appeared in the subcircuits corresponding to the skin, skin/electrode, and skin-to-electrode interface, respectively. Owing to the strong adhesion of the CNT-GO-PDMS patch on the skin, the skin and electrode surface were fused into a circuit without forming a subcircuit corresponding to the interface. These results confirmed that the amplitude and SNR of EEG were improved with the combination of the contact current of the Ag-claw and the noncontact capacitance mode of CNT-GO-PDMS, but the contribution rate of both could not be clearly quantified.

Visually Evoked Electroencephalography Signals of Steady-State Visual-Evoked Potentials Paradigm

Figure 4A shows a specific scheme of the SSVEP paradigm. The visual stimulation interface was composed of nine flashing blocks that flashed for 5 s every 1 s; the flashing blocks with the same stimulation frequency repeated the stimulation three times. The frequencies of the scintillators were composed of five α-rhythms (9.5, 10.5, 11, 11.5, and 12.5 Hz) and four β-rhythms (13, 13.5, 14, and 14.5 Hz). The scalp position for recording the EEG signals was composed of P3, Pz, and P4 channels in the parietal area and O1, OZ, and O2 channels in the occipital area (Figure 4B). The time-domain spectrum of the visually evoked EEG gradually became denser with an increase in stimulation frequency from 9.5 to 14.5 Hz (Figure 4C and Figure S15 and S16, Supporting Information). By analyzing the frequency spectra of the nine evoked EEG signals, feedback signals were observed at the frequency position of the scintillator; that is, the frequency of the EEG peak was consistent with the stimulation frequency (Figure 4D and S17, Supporting Information). The amplitude of the evoked EEG signals recorded using the Ag/CNT-GO-PDMS electrode was greater than 3 μV, confirming the excellent visually evoked EEG response (Figure 4D). Taking the frequency-domain spectrum and time–frequency diagram of the 9.5 Hz EEG signal as an example, successful acquisition of the evoked EEG signals was confirmed by observing the frequency harmonics twice, thrice, and four times (Figure 4E,F and S18, Supporting Information). Figure 4G shows the amplitude distribution of the EEG signals induced by the nine stimulus frequencies of 9.5 to 14.5 Hz. The maximum amplitude of all Ag/CNT-GO-PDMS channels ranged from 4 to 5 μV. For electrode channels, the amplitudes of the right parietal and occipital lobes were significantly higher than those of the left parietal and occipital lobes, while the amplitude in the occipital lobe was significantly higher than that in the parietal lobe. We compared the SNRs of the electrode channels in the parietal (P3, Pz, and P4) and occipital (O1, Oz, and O2) regions and found that the average SNRs in the occipital region were higher than those in the parietal region (Figure 4H). Therefore, the three channels (O1, Oz, and O2) in the occipital region were selected for the subsequent joint application of visual and auditory stimuli.

Figure 4. A) Cycle of the SSVEP offline experiment. B) Electrode location diagram. C) Visually evoked EEG signals (on Oz) under nine stimulation frequencies (9.5, 10.5, 11, 11.5, 12.5, 13, 13.5, 14, and 14.5 Hz) in the SSVEP paradigm. D) Frequency–amplitude spectra of the evoked EEG signals in (C). E) Frequency–amplitude and F) time–frequency spectra of evoked EEG signals with a stimulation frequency of 9.5 Hz. G) The amplitude distributions of the three electrode channels (O1, Oz, and O2) in the occipital region. H) Average SNR values on occipital (O1, Oz, and O2) and parietal (P3, Pz, and P4) channels.

Auditory-Evoked Electroencephalography Signals of Multiple Auditory Steady-State Response Paradigm

Figure 5A introduces the design scheme of the MASSR paradigm in the form of an experimental flowchart. The specific flow of the experimental scheme is as follows: audio files with a single frequency were produced based on the sound function in MATLAB, which were played to obtain evoked EEG signals, and then the characteristics of the feature signals were extracted to generate the frequency–amplitude spectrum of the evoked EEG signals. The temporal lobe is primarily responsible for the processing of auditory information. Therefore, four Ag/CNT-GO-PDMS electrodes were placed on the TP9, TP10, FT9, and FT10 channels in the temporal lobe (Figure 5B).

Figure 5. A) Experimental flowchart of auditory steady-state response (ASSR). B) Positions of the four electrodes in the temporal region. C) Frequency–amplitude spectra recorded using Ag/CNT-GO-PDMS and Ag/AgCl electrodes, with a stimulation frequency of 425 Hz. D) Amplitude-distribution topography recorded from the scalp of a volunteer using four Ag/CNT-GO-PDMS electrodes distributed in the temporal region, under approximately 200–1000 Hz stimulation. E) EEG SNRs of Ag/AgCl and Ag/CNT-GO-PDMS electrodes on four channels of TP9, FT9, FT10, and TP10.

In the ASSR paradigm, we recorded the evoked EEG signals using a commercial gel Ag/AgCl wet cap and self-made Ag/CNT-GO-PDMS electrodes (Figure S19–S23, Supporting Information). Nine pure tones were used for stimulation in the range of 225–1025 Hz (Audio S1–S9, Supporting Information). The frequency–amplitude spectra of the feedback EEG signals confirmed the feasibility of ASSR induction over a wide frequency range. In addition, the EEG amplitude of high-frequency sounds was lower than that of low-frequency sounds (Figure S21 and S23, Supporting Information). Figure 5C shows the amplitude spectra of evoked EEG signals at 425 Hz. The response amplitude of the Ag/CNT-GO-PDMS electrodes was higher than that of Ag/AgCl electrodes. Figure 5D shows the EEG amplitude distribution in the temporal region for nine pure-tone feedbacks. Furthermore, compared with the TP9 and TP10 channels, relatively high-amplitude characteristics were observed on the FT9 and FT10 channels. The amplitude of the evoked EEG signals ranged from 0.023 to 0.35 μV. Furthermore, we compared the SNR values of the self-made Ag/CNT-GO-PDMS electrode with those of the gel Ag/AgCl wet electrode in the four channels (Figure 5E). The SNRs of the self-made electrode were significantly higher than those of the Ag/AgCl electrodes. For feedback EEG signals with frequencies of 225, 325, 425, 525, 625, 725, 825, 925, and 1025 Hz, the average SNR of the self-made Ag/CNT-GO-PDMS electrode is 13.2 dB, and the average SNR of the Ag/AgCl wet electrode is 9.1 dB (Figure 5E).

We repeatedly recorded auditory-evoked EEG signals in the range of 225–1025 Hz, with 1500 Hz as the fundamental frequency in the MASSR paradigm (Audio S10–S18, Supporting Information). The audio files were arranged in the following order: fundamental frequency (1500 Hz), target frequency (225–1025 Hz), and fundamental frequency (1500 Hz), and the playback times were 5, 5, and 5 s, respectively. Figure S24, Supporting Information, shows the frequency domain spectra of the evoked EEG signals recorded under the stimulation of the fundamental and target frequencies. Compared with EEG signals without the fundamental frequency, the amplitude and SNR of evoked EEG signals recorded with the fundamental frequency were not significantly different (Figure S23, S25, and S26, Supporting Information). However, the semantic stimulation paradigm based on fundamental frequency reduces the efficiency of semantic decoding owing to the complexity of the data selection and feature extraction.

Audiovisual-Evoked Electroencephalography Signals of the Steady-State Visual-Evoked Potentials–Multiple Auditory Steady-State Response Paradigms

We designed the EEG acquisition experiment of “one” to “nine” speech recognition using visual–auditory stimulation, namely the SSVEP–MASSR paradigm. In the experimental scene shown in Figure 6A, a subject wears the Ag/CNT-GO-PDMS-based EEG cap and a helmet earphone and looks at the flashing squares on display (Audio S19–S27, Supporting Information). As an example, the detailed experimental scheme for identifying the word “eight” is displayed (Figure 6B). First, the audio file and 14 Hz visual stimulation were started simultaneously. In the sound-stimulation process, the pronunciation of “eight” and four audio frequencies of 212, 427, 641, and 860 Hz were played in order (Audio S25, Supporting Information). Figure 6C shows the software code fragments for the SSVEP–MASSR paradigm. The total duration of simultaneous visual–auditory stimulation was 21.1 s. The visual stimulus of the SSVEP paradigm lasted for 21.1 s. For the MASSR paradigm with a total duration of 21.1 s, the word play time was 1.1 s, and the play time of four pure tones with different frequencies was controlled at 5 s each. Figure 6D shows the scalp locations of the seven Ag/CNT-GO-PDMS electrodes. In the SSVEP–MASSR paradigm, seven self-made EEG electrodes were used, which were distributed in the occipital O1, Oz, and O2 channels, as well as the temporal TP9, FT9, FT10, and TP10 channels.

Figure 6. A) Scene depicting the SSVEP–MASSR stimulation experiment. B) Flowchart of the experiment. C) Screenshot of SSVEP–MASSR stimulating software code. D) Scalp channels of the seven Ag/CNT-GO-PDMS patches on the EEG cap. E) Sound waves of the word “eight,” visual-evoked EEG segment, and auditory-evoked EEG segments.

Figure 6E and S27–S32, Supporting Information show the sound waves of the pronunciation of the numbers one to nine. Each word was decoded into four pure-tone combinations in the frequency range of 200–1000 Hz, indicating that the pronunciation of the nine words was decoded into an EEG language (Figure S27 and Audio S19–S27, Supporting Information). The responsiveness of the Ag/CNT-GO-PDMS electrode to the feedback EEG signals of the SSVEP and ASSR paradigms was confirmed (Figure 4, 5, and 6E). EEG recognition results of nine word sounds in a separate MASSR paradigm were investigated for the first time. The decoded pure tones were used as auditory-stimulus sound waves, the evoked EEG signals of the temporal and occipital regions (FT9, FT10, TP9, TP10, O1, Oz, and O2) were recorded, and the response features were successfully extracted from the frequency-domain spectra (Figure S33–S37, Supporting Information). In addition, feedback EEG signals were recorded for comparison in a separate SSVEP paradigm (Figure S38–S43, Supporting Information).

Figure 7A shows the amplitude distribution of the visual–auditory-evoked EEG signals on the scalp within the SSVEP–MASSR paradigm. For the recognition of the words “one,” “two,” “three,” “four,” “five,” “six,” “seven,” “eight,” and “nine,” the SSVEP stimulation frequencies were 9.5, 10.5, 11, 11.5, 12.5, 13, 13.5, 14, and 14.5 Hz, respectively (Figure S44–S48, Supporting Information). In the MASSR–SSVEP paradigm, the amplitudes of the visually evoked EEG signals changed in the range of 1.1–1.8 μV (Figure S49, Supporting Information). For the pronunciations of the nine words, the dominant frequencies of their sound waves were distributed at ≈206 ± 10, 418 ± 11, 599 ± 56, and 826 ± 49 Hz (Figure 7A and S50–S58, Supporting Information). However, the four frequencies of the main peaks of the nine words were different, thus serving as a qualitative basis for identifying words. For example, the sound wave of “seven” was composed of 224, 445, 664, and 893 Hz peaks, while the word “eight” comprised 212, 427, 641, and 860 Hz peaks. The amplitude ratio of the four main peaks of the word “seven” was 2.7:1.8:1.0:1.1, while that of the word “eight” was 2.0:1.3:1.0:1.0 (Figure 7B). When more peaks were set as the qualitative basis, semantic recognition improved; that is, the pronunciation of each word was more accurately distinguished. In addition, the amplitudes of the electrode channels differed in the pronunciation of each word, serving as a quantitative basis for evaluating accuracy.

Figure 7. Amplitude distribution of evoked EEG signals in response to the words “one” to “nine” using a seven-lead Ag/CNT-GO-PDMS-based EEG cap (TP9, FT9, FT10, TP10, O1, Oz, and O2). A) Visual–auditory integration. B) MASSR-EEG language and frequency spectra of the feedback EEG signals for the sound waves of “seven” and “eight.”

However, the volunteers’ natural language was Chinese rather than English. We created Chinese speech stimulation files and repeated the MASSR experiment (Audio S28–36, Supporting Information). For the MASSR paradigm with a total duration of 16 s, the word play time was 1 s, and the play time of the three pure tones with different frequencies was controlled at 5 s each. The nine Chinese words were composed of three characteristic peaks with different frequencies, serving as the qualitative basis for word recognition (Figure S59–S61, Supporting Information). For example, the sound wave of “seven” comprised 307, 619, and 937 Hz peaks, while the word “eight” comprised 258, 517, and 775 Hz peaks. The amplitude ratio of the three main peaks of the word “seven” was 1.4:1.0:1.0, while that of the word “eight” was 2.3:1.0:1.0 (Figure S62, Supporting Information). These results further confirmed that the Ag/CNT-GO-PDMS-EEG cap-based MSSR paradigm recognized word semantics based on the EEG feature signals of word pronunciation, which was not limited by the level of language familiarity of volunteers.

Figure 8 shows the accuracy (P) of EEG recognition for nine words in the visual, auditory, and visual–auditory-evoked paradigms. According to previous research on the SSVEP–BCI system, when the amplitudes of the frequency-domain spectra are greater than 1 μV, the accuracy of online analysis reaches 87.5%.^[22] The accuracies of the SSVEP were determined based on the 1 μV amplitude of the frequency-domain spectra of the EEG signals. The classification of the MASSR was based on the test results of the gel-Ag/AgCl wet electrode. The average value of the classified amplitudes of the feedback EEG with various frequencies is ≈0.11 μV and the corresponding relationship between the specific frequencies and critical amplitudes is shown in Table S1, Supporting Information.

Figure 8. Accuracy of the 7-lead Ag/CNT-GO-PDMS-based EEG cap to recognize the evoked EEG signals of the words “one” to “nine.” In A) separate and B) synchronous stimulation, the visual evoked (I), auditory evoked (II), and their equal-proportion superposition (III). C) Statistical evaluation of the effects of stimulus frequency, EEG channel, and words on the accuracy of word recognition for separate (I) and synchronous (II) stimulation (**p [less than] 0.01 and *p [less than] 0.05).

For the individual-evoked EEG signals of the SSVEP paradigm, the three Ag/CNT-GO-PDMS patches in the occipital region (O1, O_Z, and O2) were highly accurate with an average accuracy of 89.6% (Figure 8A–I). Interestingly, the average accuracy for visual–auditory induction was 90.4%, which was higher than that for visual induction alone (Figure 8B–I). For visual–auditory induction, the average accuracy of the three Ag/AgCl electrodes in the occipital region was 74.1% (Figure S63, Supporting Information). We also investigated the accuracy distribution of auditory-evoked EEG signals (Figure 8A-II and B-II). The average accuracy in the temporal region (FT9, FT10, TP9, and TP10) was 36.7% for auditory induction and 54.0% for visual–auditory synchronization. The average accuracy of the four gel-Ag/AgCl wet electrodes was 44.9%. The average accuracy of visual and auditory-evoked EEG responses after 1:1 superposition reached 39.5%. However, the average accuracy of the EEG responses induced by visual–auditory integration was improved to 50.1%. For visual–auditory integration, the average accuracy of the gel-Ag/AgCl wet electrodes was 33.4%. For example, the total accuracy of the word “seven” was 35.7% for single induction and 50.4% for simultaneous induction; the total accuracy of the word “eight” was 41.1% for single induction and 53.2% for simultaneous induction. This improvement could be related to the sensory-level integration effect.^[23] Thus, we conducted an experiment on the brainwave transmission of words in a simulated business meeting scenario using the SSVEP–MASSR paradigm (Video S1, Supporting Information). The subjects wearing the Ag/CNT-GO/PDMS-based EEG cap only passively heard and saw the stimulus signals sent by others, but they did not know the content of the received information. However, third-party personnel successfully received the word “eight,” and the success rate of brainwave transmission reached 100% (Figure S64, Supporting Information).

Subsequently, analysis of variance (ANOVA) was used to analyze the variability of the three factors: stimulus frequency, EEG channel, and word accuracy (Figure 8C). The significance level thresholds were 0.05 and 0.01. Judging from the F-values of these three factors, the primary and secondary factors were ranked as follows: Frequency > Channel > Word. This ranking applied to both individual (F:18.8 > 4.95 > 2.55) and synchronous (34.3 > 2.3 > 0.7) stimulations. For the recognition of nine words, the stimulus frequency was a very significant factor (**p < 0.01, n = 63), which is consistent with the conclusion that frequency is the qualitative basis for word recognition. In addition, the EEG channel is a very significant factor (**p < 0.01, n = 45) in a single-stimulus experiment, whereas its significance was lower (0.01 < *p < 0.05, n = 45) in a mixed-stimulus experiment owing to the improvement in overall accuracy. In addition, the word was a significant factor (0.01 < *p < 0.05, n = 35) in the single-stimulus experiment, but was not a significant factor (p > 0.05, n = 35) in the synchronous stimulus experiment. This also shows that the new strategy based on the SSVEP–MASSR paradigm can be applied to recognize more words. In general, these results confirm that the BCI system established by connecting the self-made Ag/CNT-GO-PDMS patches with electrode channels in specific brain regions can recognize various words using the SSVEP–MASSR paradigm, representing a new strategy for establishing an EEG language database.

Conclusion

Flexible Ag/CNT-GO-PDMS-based EEG patch electrodes were prepared by combining spin-coating and heat-curing technologies. The conclusions are summarized as follows. First, we propose a cooperative EEG sensing mechanism consisting of non-contact capacitive and contact current modes to reduce scalp contact impedance and improve the EEG SNR, which is suitable for recording EEG signals with extremely low amplitudes. The CNT-GO-PDMS matrix was tightly attached to the scalp, and electrons were obtained from the scalp in non-contact capacitive mode. The internal CNT-GO conductive network quickly transferred electrons to the Ag claws, which promoted the collection of EEG signals with a high SNR. Second, we proposed a strategy for establishing EEG language. The pronunciations of the nine words (i.e., numbers 1–9) were decomposed into four pure tones, which were successfully recorded in the MASSR paradigm. For the EEG signals of the MASSR paradigm collected in the temporal region using the flexible Ag/CNT-GO-PDMS patches, the semantics were recognized according to the dominant frequency and maximum amplitude of the word pronunciation. Third, volunteers wore a 7-lead Ag/CNT-GO-PDMS-based EEG cap and collected EEG signals in the SSVEP-MASSR paradigms, demonstrating the high accuracy of word recognition. Finally, the self-made flexible Ag/CNT-GO-PDMS patches can still be classified as EEG wet electrodes because they use NaCl electrolytes or require assistance from sweat. In future research, we will optimize the distribution of the capacitance and current modes while preparing the Ag/CNT-GO-PDMS dry electrode. We will also establish an EEG semantic database based on MASSR and SSVEP–MASSR paradigms to lay the foundation for developing an online EEG semantic communication software system.

Experimental Section Preparation of Silver/Carbon Nanotube-Graphene Oxide-Polydimethylsiloxane Electrode Patches

Figure 1A illustrates the preparation process of the Ag/CNT-GO-PDMS electrode patches. A PDMS prepolymer mixture was prepared by mixing the elastomer (2.18 g) and curing agent (0.218 g) at a mass ratio of 10:1. In the PDMS prepolymer mixture, the GO (ϕ:0.2–10 μm, thickness:1 μm) and CNT (ϕ:1–2 nm, length:5–30 nm) powders were added during stirring and then ultrasonically dispersed for 5 min, thus preparing a CNT-GO-PDMS spin-coating glue. The total mass of GO and CNT powder was 100 mg, and the mass fraction of the CNT powder was adjusted to 0%, 25%, 50%, 75%, and 100%. The CNT content was 50%, unless otherwise specified.

Cu foil with a thickness of 0.02 mm was polished in a clockwise direction with sandpaper (≈68 nm) and washed with ultrapure water prior to use. The Cu foil was sucked onto the sample stage, and the CNT-GO-PDMS spin-coating glue was dropped gradually in the proportion of 1 mL of glue in the 15 mm diameter range. In specific operations, the stage was rotated at 600 rpm for 6 s, allowing the colloid to uniformly cover the surface of the Cu foil. Next, the glue was spun at a rotational speed of 1200 rpm for 10 s, and a CNT-GO-PDMS prepolymer film was deposited on the Cu foil. Finally, the spin-coating film was cured under an IR lamp (100–300 W) at 150 °C for 10 min to obtain the CNT-GO-PDMS patch (Figure S65A, Supporting Information).

A porous polyethylene template with 19 holes was prepared. The diameter of the holes was 0.45 mm, the spacing was 2 mm, and the volume was 3.18 μL. The CNT-GO-PDMS electrode was covered with the porous template, and conductive silver paste was injected into the holes. Finally, Ag/CNT-GO-PDMS electrodes were obtained by thermal curing at 150 °C for 5 min (Figure S65B,C, Supporting Information).

Assembly of Silver/Carbon Nanotube-Graphene Oxide-Polydimethylsiloxane Electrode-Based Electroencephalography Cap

The Ag/CNT-GO-PDMS-based EEG cap was composed of seven electrodes, that is, seven channels. The electrode positions in the temporal and occipital regions were sensitive to auditory- and visual-evoked EEG responses, respectively. According to the electrode positions of the 10–10 EEG standards, the electrode distributions in the temporal (TP9, FT9, FT10, and TP10) and occipital regions (O1, Oz, and O2) were determined. The electrode holders composed of the ABS resin round shell (ϕ 20 mm) and Cu metal lining (ϕ 9.4 mm) were fixed on the flexible nylon belt with a buckle hole (20 mm). Before determining the EEG signals, disposable flexible Ag/CNT-GO-PDMS films were pasted onto each electrode holder with double-sided carbon conductive tape (≤5 Ω mm⁻²), producing an Ag/CNT-GO-PDMS-based EEG cap (Figure S65D, Supporting Information).

Characterization

The morphology, composition, and chemical structure of the CNT-GO-PDMS and Ag/CNT-GO-PDMS electrode patches were characterized by field-emission SEM (FESEM; Carl Zeiss, MERLIN Compact, Germany; FEI, Verios 460L, USA), optical microscopy (OM; 6XBPC, Shoif, China), XPS (ESCALAB 250 Xi, Thermo Scientific, USA), Raman spectroscopy (LabRAM HR Evolution, Horiba Scientific, France), and FTIR spectroscopy (Tensor 27, Bruker, Germany).

Measurement of Skin-Contact Impedance

The three-electrode system consisted of an Ag/CNT-GO-PDMS working electrode, Ag/AgCl reference electrode, and Pt counter electrode, which were placed on the back and both sides of the hand. The three-electrode system was connected to an electrochemical workstation (CHI 660E, Shanghai CH Instrument Company, China), and the electrochemical impedance spectrum (EIS) was recorded using a saturated NaCl solution (10 μL) as the electrolyte. For the EIS measurements, the initial voltage was 0.5 V, the frequency range was 1–100 kHz, and the amplitude was 5 mV.

Electroencephalography Measurements

Ten volunteers signed a written informed consent form before participating in the study. All procedures carried out in the studies involving human participants were conducted in accordance with the ethical standards of the ethics committee of Tianjin University (Approval No. TJUE-2022-044), which fully complied with the ICH-GCP and the related laws and regulations of the People's Republic of China.

The EEG signals were recorded using a Neuroscan Grael EEG amplifier (Grael, Compumedics Ltd. Melbourne, Australia) at a sampling rate of 4096 Hz. For the Ag/CNT-GO-PDMS and Ag/AgCl electrodes, spontaneous EEG, EMG, and EOG signals were recorded synchronously in the paradigms of closed/open eyes, blinking eyes, and tooth grinding. The former was recorded in the occipital area with hair (Oz and Pz), and the latter two were recorded in the forehead (Fp1 and F4). Ag/AgCl electrodes secured to the ear lobes served as the reference and ground electrodes (A1 and A2, respectively). All Ag/AgCl electrodes used to measure EEG signals were coated with Quik-gel (40 μL) (Compumedics Limited). In addition, a saturated NaCl solution (10 μL) was dropped onto the surface of the Ag/CNT-GO-PDMS electrodes.

The correlation coefficient (R_cc) was calculated as follows[Image Omitted. See PDF]where R_cc is the correlation coefficient, cov (x, x) is the covariance between the two variables, and x1 and x2 are the filtered signals recorded by the Ag/CNT-GO-PDMS and Ag/AgCl electrodes, respectively.

The SNR was calculated as follows[Image Omitted. See PDF]where SNR is in dB, and y(f) (= amplitude (μV)/$\surd f(\text{Hz})$) is the amplitude spectrum calculated by the fast Fourier transform.

Spontaneous and visually evoked EEG signals were filtered using a 5–45 Hz band-pass filter. The auditory-evoked EEG signals were filtered using a 5–1050 Hz band-pass filter. Open-source EEGLAB from MATLAB toolboxes/scripts was used to extract and analyze the EEG data.

Description of the Steady-State Visual-Evoked Potentials Experiment

For the visual stimulation originating from the square flickering block (2.6 cm × 2.6 cm), the stimulation frequencies were set to 9.5, 10.5, 11, 11.5, 12.5, 13, 13.5, 14, and 14.5 Hz. The visual stimulator was a liquid-crystal display with a resolution of 1920 × 1080 pixels and a refresh rate of 60 Hz. The test environment was a room without windows and with normal lighting. The distance between the volunteer's chair and the stimulator was 60 cm. For each experiment, the stimulation lasted 5 s and was repeated thrice, with a 1 s interval between each stimulation.

The SNR was calculated as follows[Image Omitted. See PDF]where y(f) (= amplitude (μV)/$\surd f(\text{Hz})$) is the amplitude spectrum calculated using fast Fourier transform.

Description of Auditory Steady-State Response, Multiple Auditory Steady-State Response, and Steady-State Visual-Evoked Potentials–Multiple Auditory Steady-State Response Experiments

Voice files of the words “one” to “nine” were edited in the Adobe Audition CC2019 version. We produced 36 voice files, including nine pure-tone files of 225, 325, 425, 525, 625, 725, 825, 925, and 1025 Hz (Audio S1–S18, Supporting Information) and soundwave files of the nine words (Audio S19–S36, Supporting Information). The auditory stimulator was a pair of SONY MDR-XB950B1 headphones. In the ASSR paradigm, the stimulation frequencies were set to 225, 325, 425, 525, 625, 725, 825, 925, and 1025 Hz for 10 s each time, with a sound intensity of 30 dB. The volunteers wore a homemade EEG cap and headphones, and various frequency stimuli were played using MATLAB software. The stimulus tone of each frequency was played once, and the duration of each stimulus was 10 s without loop playing. When recording the evoked EEG signals of the MASSR paradigm, the volunteers looked at a static white square (4 × 4 cm) on the display screen with a flashing frequency of 60 Hz.

In the semantic stimulus files of the nine words, each file was spliced with the English pronunciation of the Arabic numeral and the corresponding four stimulus frequencies. MATLAB 2016a software (Psychology Software Tools, Inc., USA) was used for auditory stimulation (playing audio), and Curry 8.0 (Compumedics Neuroscan, Australia) was used to collect EEG signals synchronously. A volunteer wore headphones and auditory-feedback EEG signals were recorded as the voice files were played. In addition, synchronous-feedback EEG signals for vision and hearing were collected through simultaneous SSVEP stimulation. The SNR was calculated according to Equation (2).

For the classification, a method of directly comparing the amplitude of the measured EEG data with the critical amplitude value was used, and the amplitude was obtained from the frequency spectrum. Accuracy (P) was defined as the percentage of effective times over the total number of experiments. To clearly express the statistical method of P, the following formula was used[Image Omitted. See PDF]where N_eff is the number of times the amplitude is greater than the critical value in each person's repeated experiment, N_E and N_p are the number of electrode channels and the number of volunteers, respectively, and N_T is the total number of experiments.

The experiment involved nine words, from “one” to “nine,” and one word was composed of five frequencies. The research included ten volunteers, and each of them repeated the experiment five times. More than 70 electrodes were prepared. The ten volunteers used the same EEG-cap for the same experimental item, which was beneficial for eliminating the influence of the electrode and ohmic contact. The statistical results after introducing the number of experiment repeats (n = 5) and the number of people (n = 10) are as follows:

For frequency: n[channel] × n[word] × 50 = 7 × 9 × 50 = 3150.

For the EEG channel: n[word] × n[frequency] × 0 = 9 × 5 × 50 = 2250.

For word: n[channel] × n[frequency] × 50 = 7 × 5 × 50 = 1750.

Acknowledgements

P.L. and C.W. contributed equally to this work. This work was financially supported by the National Nature Science Foundation of China (No. 62271350) and the Natural Science Foundation of Tianjin City (Nos. 20JCZDJC00290 and 22YDTPJC00650).

Conflict of Interest

The authors declare no conflict of interest.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.