research-article

Open access

WalkingWizard—A Truly Wearable EEG Headset for Everyday Use

Authors:

Teck Lun Goh,

Li-Shiuan PehAuthors Info & Claims

ACM Transactions on Computing for Healthcare, Volume 5, Issue 2

Article No.: 9, Pages 1 - 38

https://doi.org/10.1145/3648106

Published: 22 April 2024 Publication History

PDF eReader

Abstract

Electroencephalography (EEG) provides an opportunity to gain insights to electrocortical activity without the need for invasive technology. While increasingly used in various application areas, EEG headsets tend to be suited only to a laboratory environment due to the long preparation time to don the headset and the need for users to remain stationary. We present our design of a dry, dual-electrodes flexible PCB assembly that realizes accurate sensing in the face of practical motion artifacts. Using it, we present WalkingWizard, our prototype dry-electrode EEG baseball cap that can be used under motion in everyday scenarios. We first evaluated its hardware performance by comparing its electrode-scalp impedance and ability to capture alpha rhythm against both wet EEG and commercially available dry EEG headsets. We then tested WalkingWizard using steady-state visual evoked potential (SSVEP) experiments, achieving high classification accuracy of 87% for walking speeds up to 5.0 km/h, beating state-of-the-art. Expanding on WalkingWizard, we integrated all necessary electronic components into a flexible PCB assembly—realizing WalkingWizard Integrated, in a truly wearable form-factor. Utilizing WalkingWizard Integrated, we demonstrated several applications as proof-of-concept: classification of SSVEP in VR environment while walking, real-time acquisition of emotional state of users while moving around the neighbourhood, and understanding the effect of guided meditation for relaxation.

1 Introduction

Over the years, EEG applications have evolved from primarily clinical environments such as the diagnosis of irreversible brain death [68], epilepsy [5], and the use of brain-computer interface (BCI) for neuro-rehabilitation [7, 16, 65, 73], to the consumer market with applications such as neuromarketing [13, 19, 31] and gaming [46]. In recent years, EEG has also been used to study social interactions [53] and for human factors engineering [26, 35]. Market research expects the EEG device market to expand at an annual growth rate of 8.7% [2], with diverse applications of EEG becoming widespread in our everyday lives.

For EEG devices to be pervasive, the wearability of the EEG device is critical. Users should be able to put on/take off the EEG device easily. This points to the need for fully dry electrodes, as wet [43–45, 55, 69, 70, 73, 76] and semi-dry electrodes [49] require long preparation time prior to wearing [11], periodic checks of the electrode impedance and replenishing of the saline [39], not to mention the need for external assistance to don the device [39, 75].

Also, they need to work well under real-life scenarios, where the user is in natural motion (including while in fast motion) going about his everyday activities [35], instead of being constrained in laboratory experiments to remain stationary or be restricted to a limited set of movements [19]. However, as motion artifacts heavily corrupt the EEG recording, especially in dry electrode EEG devices [71], our EEG device needs to be resistant to motion artifacts despite its use of dry electrode EEG.

EEG devices with many electrodes will add to the bulk of the wearable, limiting their use to the laboratory environment [39], so a truly wearable EEG device should have a low electrode count. Next, the EEG device should be able to sense and respond in real-time, keeping up with the natural pace of the user without being laggy. Finally, the EEG device should be aesthetically unobtrusive [24] and portable to allow users to go about his everyday activities.

As we will detail in Section 2, no off-the-shelf EEG headset satisfies all the above requirements (see Table 1). To achieve our vision of a truly wearable EEG headset for everyday use, we conceived the idea of a dry, dual-electrode flexible PCB assembly—a lightweight, modular flexible PCB assembly that can be fixed to the underside of any cap for EEG recording during motion. This forms the basis of our WalkingWizard prototype—a EEG headset built atop a baseball cap, that works under natural motion for everyday use in real-time. The accompanying software processing pipeline of WalkingWizard then pre-process the recorded EEG signals and feeds it into the state-of-the-art EEG deep learning model (EEGNet) to classify the EEG signal to its respective steady-state visual evoked potential (SSVEP) stimuli.

Table 1.

Related Work	Resistant to Motion Artifacts	Fully-Dry Electrodes	Electrode Count	Real-time	Aesthetics
[14, 45, 63, 64, 69]	Yes	No	32-256	No	Tethered, Obtrusive
[70, 73]	Yes	No	64-128	Yes	Tethered, Obtrusive
[55, 76]	Yes	No	64	Yes	Obtrusive
[43, 44]	Yes	No	32	Not Indicated	Obtrusive
[19, 49]	Yes	Semi-Dry	14	Yes	Yes
[22] - actiCAP Xpress	No	Yes	16	No	Obtrusive
[22] - Trilobite	No	Yes	32	No	Obtrusive
[22] - Starstim	No	Yes	8	No	Obtrusive
[22] - JellyFish	No	Yes	4	No	Yes
[47, 75]	Min. movement	Yes	8	No	Yes
[25]	Min. movement	Yes	6	Yes	Obtrusive
[48, 50]	Yes	Yes	8	No	Obtrusive
[58, 59]	Yes	Yes	64	Yes	Obtrusive
WalkingWizard	Yes	Yes	5	Yes	Yes

Table 1. Summary of Related Works in EEG Headsets

We recruited eight volunteers to perform SSVEP experiments under various motion states to compare the performance of WalkingWizard against state-of-the-art.

We further developed WalkingWizard Integrated—a truly wearable dry electrode-based EEG headset with the accompanying acquisition electronics integrated into a single flexible PCB assembly. With WalkingWizard Integrated, we demonstrated several prototype applications for everyday situations such as Virtual Reality (VR) settings, daily commuting, and while performing guided meditation sessions. Overall, this article made the following contributions:

—

Contribution 1: First in literature to build a lightweight, modular, dry, dual-electrode flexible PCB, assembled with spring-loaded pins for EEG recording under motion.

—

Contribution 2: Integrated the various hardware components with our software implementation to realize the WalkingWizard prototype that can work under natural motion.

—

Contribution 3: Performed and compared the hardware performance of WalkingWizard against gold-standard wet electrode EEG and other dry EEG headsets, demonstrating the ability of WalkingWizard to perform EEG recording despite presence of motion artifacts.

—

Contribution 4: Performed user trials on SSVEP experiment to verify the performance of WalkingWizard under different walking speeds, achieving the highest classification accuracy of 87% versus the state-of-the-art.

—

Contribution 5: Integrated all necessary electronic components into a flexible PCB assembly and developed WalkingWizard Integrated, a truly wearable dry EEG headset that works under natural motion.

—

Contribution 6: Demonstrated several applications: (1) robust and real-time EEG signal acquisition in a VR setting while walking, (2) a prototype wellness application to record and display the real-time emotional state (Frontal Alpha Asymmetry (FAA)) of users while cycling and walking around the neighbourhood, and (3) emotional state monitoring while performing guided meditations.

2 Related Works

We classify the prior related works in EEG headsets based on the type of electrodes below.

EEG headsets with wet electrodes. References [43–45, 55, 70, 73, 76] use wet electrodes to capture EEG signals during motion due to their superior performance in signal integrity and ability to tolerate motion artifacts [18, 71]. Ferris et al. introduced the dual-electrode hardware configuration [62, 69] and demonstrated effective motion artifact removal with high density wet electrode EEG headset [14, 63, 64]. However, the long preparation time involved in putting on wet electrodes, and the associated discomfort, limits the use of headsets with wet electrodes to laboratory settings.

EEG headsets with semi-dry electrodes. Emotiv Epoch+ was used in References [19, 49] as they provide a compromise between the amount of setup preparation required and resistance to motion artifacts. However, users are required to add saline solution to the sponge electrode periodically (approximately every 30 min [49]) to maintain the electrode impedance, making semi-dry electrodes still not practical as a wearable for everyday use. In Reference [19], while EEG is recorded throughout the natural motion of the users, EEG signals were analyzed only during the time interval when they were not moving.

EEG headsets with dry electrodes. Dry electrodes are the most practical for wearability. However, motion artifacts are a challenge. Several prior works were able to handle limited user head movement, such as when the person is seated in flight [25], performing stationary cycling exercises [47], and walking around a art museum [22]. In Reference [75], users walked slowly (2.0 km/h) on a treadmill, while the auditory steady-state response (ASSR) was recorded. The experiments in Reference [75] were designed such that the slow moving speed, coupled with ASSR at 40 Hz where the spectral content attributed to motion artifact is minimum, bypassing the challenges posed by dry-electrode EEG in motion, instead of directly addressing the problem.

The first use of a dry EEG headset on a fast moving user (up to 4.8 km/h) was demonstrated in References [48, 50]. However, the computation was not done in real-time as human intervention was required to manually remove transient artifacts and noisy channels.

In Reference [58, 59], the team demonstrated real-time motion artifact removal on a 64 channel dry-electrode EEG headset with the introduction of artifact subspace removal (ASR) pre-processing algorithm. The ASR algorithm extracted statistical features from a baseline clean EEG recorded while the user remained motionless, then applied that to EEG signals that were contaminated with artifacts [63]. It assumes that EEG recorded at rest is similar to that while performing the assigned tasks, which may not be the case.

While the above [48, 50, 58, 59] have dry electrode EEG headsets that can handle significant motion artifacts, they are highly obtrusive (see Figures 1(A)–1(C)) as wearables. In comparison, we included in Figure 1(D) the Mindo-4S JellyFish in Reference [22], which is unobtrusive and can be practically worn as a user goes about his everyday activities, but cannot handle significant motion.

Fig. 1.

Table 1 summarizes the literature review based on their resistance to motion artifacts, type of electrodes (dry/wet), electrode count, real-time response and aesthetics (whether portable and unobtrusive). Through our review, we see that current state-of-the-art EEG headsets do not support a truly wearable EEG for everyday use—motivating our development of WalkingWizard.

3 System Design and Implementation

3.1 Electrode Pin

Prior research on dry EEG electrodes falls under three major categories [42]: (1) microelectromechanical system (MEMS)-based electrodes—an array of microneedles penetrating tens of \(\mu\)m through the scalp, which comes with considerable infection risks, (2) capacitive electrodes, which are highly susceptible to motion artifacts, and (3) combed-shape electrodes to penetrate the hair layer. The last approach is adopted in WalkingWizard considering its ability to work on scalp with hair, it is relative maturity [32, 33, 60, 61], and the absence of infection risk.

In WalkingWizard, we used a gold-plated surface mount spring-loaded dry electrode pin (part number: 0871-0-57-20-82-14-11-0) from Mill-Max [56] to fabricate our combed-shape electrodes. Spring-loaded electrode pins are employed in WalkingWizard to minimize fluctuations in electrode-scalp impedance. Unless the maximum stroke of the pin is exceeded when the user is in motion, the noise induced by motion artifacts can be significantly reduced. The chosen electrode pins offer four desirable features: (a) It is of length 10 mm, which is sufficient to penetrate the hair of most users [60]; (b) it is gold-plated for biocompatibility; (c) it has a stroke length of 2 mm (to minimize fluctuation in electrode-scalp impedance); (d) it offers the largest diameter (for comfort) of all the commercially available spring-loaded pins.

Four pins were chosen for each electrode as a compromise between achieving low electrode-scalp impedance (require high number of pins) and improving wearability when donning the headset (require low number of pins). To ease our flexible PCB design, these four spring-loaded pins were mounted in a symmetric square shape with 8 mm separation between pins, to form the combed-shape electrode for each electrode position. The combination of spring-loaded pins with a flexible PCB in WalkingWizard enables tight electrode contact in a comfortable form factor.

3.2 Number of Electrodes

While Reference [29] recommended more than 64 electrodes for high intensity locomotion, the high number of electrodes would make WalkingWizard bulky and difficult to wear. Using the BETA database [51]—consisting of 70 subjects performing a 40-target SSVEP experiment using a 64-channel wet electrode without motion, as benchmark, we evaluated the performance of 1, 2, 4, 8, 16, 32, and 64 electrodes in classifying the EEG signal to the SSVEP stimulus. As SSVEP responses are most readily detected in the occipital and parietal region, we selected the electrodes for analysis in the following order: occipital, parietal, temporal then frontal region. The selected electrode positions (based on the 10–10 system) for 1, 2, and 4 electrodes are provided in Table 2.

Table 2.

Number of electrodes	Electrode position
1	Oz
2	O1, O2
4	Oz, O1, O2, POz

Table 2. Electrode Positions Used to Evaluate BETA Database

Data for the first 10 subjects from the BETA database were used. The data was filtered to select only 4 classes (out of 40), and trained using EEGNet (with the default parameters of Reference [41]) after a 80%–20% train-validation split. The SSVEP classification accuracy against number of electrodes is shown in Figure 2(A).

Fig. 2.

The validation accuracy for 1 and 2 electrodes were too low to be useful. We chose four electrodes—at positions Oz, O1, O2, and POz, in our design for WalkingWizard, as it provided sufficiently high accuracy while keeping the electrode count low.

3.3 Noise Electrode

Using a silicon cap to insulate electrocortical signals from the user, [40] demonstrated that movement induced artifacts could be recorded using a high density wet electrode EEG headset placed in contact with a conductive layer above the silicon cap. Using a phantom head, Ferris et al. subsequently introduced and demonstrated the effectiveness of the dual-electrode configuration [62]—noise electrode stacked on top of the EEG electrode, in cleaning the EEG signal that has been corrupted with noise artifacts. The dual-electrode configuration was used in many other applications [14, 63, 64] on a high density wet electrode EEG headset. However, they have not been investigated for dry electrode EEG, which brings about challenges of motion artifacts.

Using 250 Bloom gelatin (mixed with 5% salt by mass based on Reference [30]) and EEG simulator (SEEG100E, WhaleTeq Co. Ltd., EEG Performance Tester), we tested our proxy dual-electrode that was designed by stacking insulative masking tape (proxy to silicon cap in Reference [40]), aluminium foil (proxy to conductive fabric) and two conductive tape as shown in Figure 3.

Fig. 3.

Conductive stretch fabric (4900 Stretch Conductive Fabric, Holland Shielding Systems BV, Dordrecht, Netherlands) used in Reference [62] was also acquired to build the set of four noise electrodes—overlaying the electrode positions as detailed in Section 3.2, shown in Figure 2(B). Black insulating masking tape was taped on the underside of the setup in Figure 2(B) to make sure that it is insulated from the EEG electrodes that it would be placed on top of.

Tests performed on both setups while in motion—shaking the setup for the former and walking on a treadmill for the latter, indicated that the noise electrode was able to detect the motion artifacts without the EEG signal. This motivated our inclusion of dual-electrode configuration into WalkingWizard.

3.4 Dry Dual-Electrode Flexible PCB Assembly

While effective in removing motion artifact, the proposed dual-electrode configuration in Reference [62] as-is is not suitable as a truly wearable EEG headset due to (1) the use of wet electrode—resulting in large amount of time required to prepare the user before and after using the headset, (2) that the headset is tethered, and (3) obtrusive form factor of the headset.

This led us to propose and design the dry, dual-electrode flexible PCB assembly that retains the ability to remove motion artifact, but in a light-weight, modular form-factor suitable for wearables. Figure 4 illustrates our flexible PCB assembly design in WalkingWizard. We condensed the overlaid conductive cap, noise electrode, and the electrical isolation between the noise and EEG electrode in Reference [62] into a single flexible PCB. One of the four layers in the flexible PCB was designed to be the noise-ground plane, with the other three layers used for signal routing. The noise-ground plane emulates the overlaid conductive cap in Reference [62], on which the noise electrodes are placed. Since all signal and plane layers in a PCB are isolated by dielectric, the dielectric layer in the PCB provide the same electrical isolation as that between the noise and EEG electrode. Signal from the noise-ground plane is tapped out as noise electrode channel. The PCB was designed on a flexible substrate using polyimide to conform to the curvature of the head when worn. To improve the reliability of the flexible PCB, stiffeners were placed on the underside of the connectors.

Fig. 4.

To replace the wet electrode with its dry equivalent, we use spring-loaded pins to improve the contact efficiencies of the electrode-scalp interface [42, 61]—thereby reducing electrode-scalp impedance. Four gold-plated surface mount spring-loaded pin from Mill-Max were assembled as combed-shaped electrodes for each of the Oz, POz, O1, and O2 electrode positions on the underside of the flexible PCB assembly, as illustrated in Figure 5. The presence of spring in the pin also minimize fluctuation in electrode-scalp impedance—as impedance remained unchanged as long as the maximum stroke of pin is not exceeded, thereby reducing induced noise due to motion artifacts.

Fig. 5.

With this flexible PCB assembly, we retained the noise removal ability of Reference [62] while eliminating the need to apply conductive gel to each electrode position, and the need to overlay a conductive cap over the set of EEG electrodes.

3.5 Analogue Filters, ADC, Wireless Transceivers

The eight-channel g.NAUTILUS Multipurpose (g.tec medical engineering GmbH) was acquired to supplement the design of WalkingWizard. In WalkingWizard, our primary focus is to design a dry-electrode EEG headset that can be used while in motion. The backend of the EEG headset—including the analogue filters, ADCs, compute module, and wireless transceivers—were common modules found in many commercial EEG headsets. Other EEG amplifiers with detachable electrodes (e.g., LiveAmp [3], Smarting mobi EEG Amplifier [6], or Cyton Board [4]) could be used in-place of g.NAUTILUS Multipurpose.

We tailored the band-pass filter of the g.NAUTILLUS Multipurpose headset to 5 to 30 Hz—to include the full spectral range of our SSVEP stimuli, up to their 3rd harmonics, while filtering away other bands that do not contain our signal of interest. The notch filter was set to 48 to 52 Hz to remove 50 Hz line noise. The sampling rate was set to 250 Hz.

3.6 WalkingWizard Overview

The four EEG channels and the noise electrode channel of the flexible PCB assembly were connected to channels 1 to 4 and channel 5 of g.NAUTILUS Multipurpose, respectively. For the ground electrode—which would be placed over the forehead (no hair) in our experiment, we used fabric conductive tape (Laird Technologies Part No. 46W5E03020.NN00). A single electrode configuration of the flexible PCB assembly was used as the reference electrode.

We mounted our flexible PCB assembly and the g.NAUTILUS Multipurpose on two different caps (used interchangeably) for our user trials, illustrated in Figure 6 as WalkingWizard versions 1 and 2. In WalkingWizard version 2, the wireless transceiver of the g.NAUTILUS Multipurpose was attached to the bottom of the tongue of the baseball cap. We employed the help of a leather seamstress to sew an additional layer of cloth over the tongue of the cap to further conceal the wires and wireless transceiver.

Fig. 6.

3.7 Software Design

3.7.1 Pre-processing.

The first software processing step would be to preprocess the signal to remove corrupted channels and/or trials that would adversely affect the subsequent processing. This could be done manually through visual inspection [48, 50], by statistical methods, or with recently proposed Artifact Components Removal (ASR) algorithm [20]. In view of the need for real-time processing, we adopted a gross artifact rejection method—similar to Reference [22], to reject channels above a certain root-mean-square (RMS) amplitude threshold.

To automatically remove corrupted channels and/or trials, the EEG recordings were first tested for correlation between the EEG channel and its adjacent channels [20], with correlation threshold set to 0.75. Channels that failed the correlation test and have RMS amplitude above 21\(\mu\)V (chosen empirically) were flagged as noisy and set to zero. Trials that contained more than two noisy channels were flagged as corrupted trials and removed.

The EEG recording was then down-sampled from 250 to 62.5 Hz to reduce the data size for subsequent steps.

3.7.2 Classification.

EEGNet [41], a compact CNN architecture, was used as the classifier for WalkingWizard. From the numerous deep-learning models for EEG classification, we chose EEGNet for our evaluations due to the promises of EEGNet being compact and its robustness across different BCI paradigms [21].

As WalkingWizard was designed with a novel hardware architecture—EEG with noise electrode channels, distinct from prevailing literature, it became imperative to fine-tune the hyperparameters of EEGNet to ensure optimal performance for our classification task.

To systematically address this requirement, we introduced an automated comprehensive framework built atop EEGNet. This framework facilitates exhaustive hyperparameter exploration, encompassing updates to EEGNet to incorporate regularization effects. The design of our framework was grounded in the understanding that achieving the best performance would require more than just adopting an existing model; it would necessitate its adaptation and evolution in harmony with our unique hardware setup.

A salient feature of our automated framework is its ability to methodically consolidate outputs from various hyperparameter tuning sessions. These outputs are compiled into a file, sorted based on the validation accuracy. It is through this iterative and organized process that we were able to discern the most suitable hyperparameter configuration, which we present below.

We provide as input a total of 92 combinations of hyperparameters for tuning. After 7 hours of hyperparameter tuning on a NVIDIA RTX 2070 8GB, the optimum set of hyperparameters for EEGNet was obtained, with:

(1) F1 = 12, (2) D = 2, (3) F2 = 24, (4) kernelLength = 256, and (5) dropout = 0.25. No regularization of EEGNet was required. Model training was done offline, but inference was carried out in real-time.

The architecture of EEGNet and its implementation in WalkingWizard is provided in Table 3.

Table 3.

Layer (Keras API [28])	No. Filters	Size	Settings
Conv2D	12	(1,256)	padding = “same”; use_bias = False
BatchNormalization
DepthwiseConv2D	2*12	(5,1)	depth_multiplier = D; use_bias = False; max_norm = 1.0
BatchNormalization
Activation			“elu”
AveragePooling2D		(1,8)
Dropout			0.25
SeparableConv2D	24	(1,16)
BatchNormalization
Activation (ELU)			“elu”
AveragePooling2D		(1,8)
Dropout			0.25
Flatten
Dense			kernel_constraint = max_norm(0.25)
Activation			“softmax”

Table 3. EEGNet Architecture and Settings Used in WalkingWizard

N-fold cross validation was adopted to evaluate the model performance for WalkingWizard. For each participant to be evaluated, the model was trained using EEG data from the other participants, split 80%–20% as training and validation set, with the data for the participant to be evaluated kept as the test set. Adam optimizer was used to fit the model to minimize the categorical cross-entropy loss function.

The EEG recording for each trial were of length 7 s. The first 1.5 s and last 0.5 s of the recorded EEG data were discarded as they captured transients that occurred before and after the trial. For the remaining 5 s of EEG recording for each trial, a sliding window of 3 s (with 98.5% overlap) was used to augment the EEG data to (1) increase the amount of data available for training, and (2) prevent the neural model from overfitting to the start phase of the SSVEP stimuli.

4 Hardware Performance

We compare the hardware performance of WalkingWizard with both the gold-standard wet electrode and also with other dry EEG headsets.

4.1 Performance Against Gold-standard Wet EEG Electrode

We compare the performance of WalkingWizard with gold-standard wet EEG electrode, without skin preparation (i.e., we did not use abrasive skin preparation gel to prepare the scalp to improve electrical contact). Our flexible PCB assembly was used to record the EEG signal from Oz, O1, and O2 electrode position. Two gold-cup electrodes were used to record the EEG signal at the electrode position between Oz-O1 and Oz-O2, respectively. The respective electrode positions are illustrated in Figure 7. “Ten20 Conductive Electrode Paste” from Weaver was used as the conductive paste to hold the gold-cup electrodes in place. Kendall H124SG self-adhesive electrode was used on the left and right mastoid as the Reference and Ground electrode, respectively.

Fig. 7.

4.1.1 Impedance.

EEG electrode-scalp impedance is measured prior to a EEG recording session to access the quality of electrode-scalp contact, which would directly impact the quality of the EEG recording. Lower electrode-scalp impedance results in EEG recording with less amount of artifact. As a reference, the guidelines for clinical grade EEG recording with skin-preparation requires scalp-electrode impedance of less than 5k\(\Omega\) [1].

A set of 12 impedance measurement was recorded once every minute. For the wet electrode without skin preparation, the electrode-skin impedance mostly (25 to 75 percentile) ranges from 31 to 45 k\(\Omega\). For WalkingWizard dry-electrodes, the impedance ranges from 174 to 231 k\(\Omega\). The impedance measurement for the wet electrode is compared against that for WalkingWizard in Figure 8. We have also included in Figure 8 the impedance measurements of the OpenBCI Cyton with dry electrodes, which will be described in more detail in the Section 4.2.1.

Fig. 8.

While WalkingWizard exhibits electrode-scalp impedance of in the range of 200 k\(\Omega\)—higher than that of wet electrode, the results are expected, as the absence of conductive gel, and combed structure of the electrodes (for hair penetration) significantly reduced the electrode-scalp contact area, thereby greatly increasing the impedance as compared to the wet electrode equivalent.

4.1.2 Alpha Rhythm Capture.

We compare the performance of WalkingWizard with wet electrode in recording the alpha rhythm from the Oz electrode position when the user’s eyes are closed. One minute of EEG recording was performed on two users with both eyes opened, and another 1 min each with both eyes closed. We performed an arithmetic mean of the two wet EEG electrodes as a proxy to the wet EEG electrode measured at Oz. We then compare this results with the actual Oz EEG data recorded by WalkingWizard. Bandpass filter settings were all set to 5 to 30 Hz. The results are provided in Figure 9.

Fig. 9.

Figures 9(A) and 9(B) shows the amplitude and spectrum plot of the EEG recording at Oz electrode position when both eyes were opened. For the amplitude plot, only the signal from the 18s to 20s of user 1 is shown for illustration. The spectrum plot provides an overview of the frequency content for the entire experiment in 0.1 Hz bin-width resolution. The results for eyes closed are shown in Figures 9(C) and 9(D). Obvious alpha rhythm in the 9.2 to 10.2 Hz region can be observed in Figure 9(D). The spectrogram plot is also provided in Figure 25 in Appendix A.1.

Fig. 10.

Fig. 11.

Fig. 12.

Fig. 13.

Fig. 14.

Fig. 15.

Fig. 16.

Fig. 17.

Fig. 18.

Fig. 19.

Fig. 20.

Fig. 21.

Fig. 22.

Fig. 23.

Fig. 24.

Fig. 25.

Pearsons correlation was used to calculate the correlation between the wet electrode and WalkingWizard for the 1min of EEG recording for both users, both when the eyes are opened and closed. The correlation is at 95% and 87% for eyes open and eyes closed, respectively.

4.2 Performance Against Other Dry EEG Headset

We further evaluate the hardware performance of WalkingWizard against several dry EEG headsets that we could obtain. The list of dry EEG headsets with their sampling rate are provided in Table 4. We performed digital bandpass filter of 5 to 30 Hz and notch filter at 50 Hz for all the recorded signal.

Table 4.

Label	Dry EEG Headset	Sampling Rate
A	WalkingWizard	250 Hz
B	Emotiv Insight 2	128 Hz
C	gtec gNAUTILUS with gSAHARA electrodes	250 Hz
D	OpenBCI Cyton with dry electrodes from OpenBCI	250 Hz
E	Emotiv MN8	128 Hz
F	OpenBCI Cyton with ThinkPulse active electrode	250 Hz

Table 4. Comparing Performance of Dry EEG Headsets with WalkingWizard

4.2.1 Impedance.

We compare the impedance performance for WalkingWizard against other dry EEG headsets.

Emotiv Insight 2 and Emotiv MN8 do not provide impedance measurement. Instead, it denotes the quality of the electrode contact using red, orange and green. Contact quality of green was obtained for our recording. gNAUTILUS also does not provide impedance measurement when interfaced with gSAHARA electrode. It is however estimated to be 208 k\(\Omega\) based on the findings from Reference [52]. While OpenBCI Cyton Board supports impedance measurement, this function is not available when interfaced with the ThinkPulse dry active electrode.

The impedance of WalkingWizard against OpenBCI Cyton Board with dry electrodes from OpenBCI is provided in Figure 8.

The impedance of OpenBCI Cyton with dry electrode ranges from 375 to 434 k\(\Omega\), higher than the 174 to 231 k\(\Omega\) of WalkingWizard.

It is to be noted that the impedance of WalkingWizard is either the same or lower than the state-of-the-art dry combed EEG electrodes in Reference [52], and it also falls on the lower range of the “few megohms to hundreds of kilohms” expected of dry EEG electrodes as published in Reference [59]. We only compare the impedance of WalkingWizard against other dry combed EEG electrodes in Reference [52] as dry flat EEG electrodes that are in contact with non-hair area have much higher surface area—resulting in much lower impedance, but are not suitable for area of the scalp with hair.

4.2.2 Alpha Rhythm Capture.

To compare the performance of WalkingWizard against other dry EEG headset, we recruited two users, and for each user recorded 1 min of EEG recording on the Oz electrode position when the user has both eyes opened, as well as both eyes are closed, in a static scenario. These experiments were then repeated on an idling vehicle (with engine turned on) whose vibration induces motion artifacts on our EEG recording.

The characteristics of the vehicle used in this experiment are provided:

—

Vehicle Model: Nissan Cabstar

—

Vehicle Year of Use: 12 years

—

Revolution per minute at Idle: 500

—

Seat Location: Driver Seat

Since there is no Oz electrode in Emotiv Insight 2, we tilted the headset when it is worn so that the Pz electrode of Emotiv Insight 2 is positioned at the Oz position. As Emotiv MN8 comes in a earbud form-factor, the EEG signal is recorded from the T7 position instead.

Figure 10 compares the performance of the six dry EEG headsets when the users have both eyes opened / closed. The alpha rhythm in the 9 to 10 Hz frequency range is apparent when the eyes are closed for all the headsets except Emotiv MN8. This is expected as Emotiv MN8 comes in a earbud form-factor and alpha rhythm when eyes are closed are not as prominent in the T7 electrode position, as compared to the Oz electrode position.

Figure 11 captures the same data when the vehicle is vibrating in its idle mode. From the figure, it is clear that while WalkingWizard continues to capture the EEG data, the high amount of noise artifact that has been induced in the other five dry EEG headset results in the alpha rhythm being lost in the spectrum plot.

We have also added in Appendix A.2 additional figures to illustrates the spectrogram of the six devices while static and in the vibrating vehicle, respectively.

5 Evaluation

We conducted user trials using SSVEP experiment to evaluate the performance of WalkingWizard across different walking speeds. The results were then compared with other state-of-the-art dry-electrode EEG headsets that performed SSVEP experiments during motion. SSVEP was used to evaluate the performance of WalkingWizard due to (1) the availability of ground-truth, and (2) the availability of published literature to compare our results with. In these SSVEP experiments, we thus kept to a maximum walking speed of 5 km/h as that is the fastest demonstrated by prior work.

5.1 Participants

Eight (six male, two female; ages 21–37) healthy participants with either normal or corrected-to-normal vision were recruited for the experiment, which was approved by the university’s Institutional Review Board. Participants were of Chinese (75%) and Indian (25%) ethnicities. All participants read and signed an informed consent prior to the conduct of the experiment.

5.2 EEG Data Acquisition

WalkingWizard was used to record the EEG data. The reference electrode—using the single electrode combed flexible PCB, was placed on the T4 electrode position, while the ground electrode—using the conductive fabric, was placed horizontally above the eye-brow. These electrode positions were chosen so that all EEG electrodes remain hidden under the cap—unobtrusive form-factor for WalkingWizard. The sample rate for WalkingWizard was set to 250 Hz.

5.3 Experiment Protocol and Procedure

In the SSVEP experiment, four visual stimuli flickering at constant but distinct frequencies were presented to the participant. The participant was instructed to shift the gaze to 1 of the 4 visual stimuli. Through the shift in gaze, the EEG signal—with frequency in the fundamental or harmonics of the flickering stimulus, could be detected at the visual cortex region (occipital lobe) of the participant.

The four visual stimuli were displayed on the four corners of a LED monitor screen (ASUS VP247HA, Refresh Rate: 60 Hz). The flickering frequency of the four stimuli were at 6.0, 6.67, 7.5, and 8.57 Hz (synchronized to the frame-rate of 60 Hz). The stimuli were designed with a black dot in the middle of each stimulus to facilitate the participant in maintaining visual focus—especially while walking. Before the stimuli started flickering, there was a prompter screen that would select (at random) and instruct the participant which of the four visual stimuli to gaze at. The participant would press the space bar to start the visual stimuli, which would flicker for 7 s. That concludes the end of one trial. The trial would be repeated 20 times to complete one run of the experiment. The stimuli for the SSVEP experiment were coded in python using the psychoPy library [67].

Participants of the user trial were requested to perform the abovementioned run of the SSVEP experiment while standing still, and while walking at four assigned speeds of 1.5, 3.0, 4.5, and 5.0 km/h on the treadmill (Xiaomi Kingsmith A1 WalkingPad Treadmill). For every movement state (and speed), the participant was requested to perform the run of the SSVEP experiment twice, followed by a short 10 min break. After the break following the end of the 5.0 km/h experiment run, the participant was requested to select and perform another two more runs of the SSVEP experiment from the following set of walking speed: 2.0, 3.5, and 4.0 km/h. The last two runs were included to evaluate the ability of WalkingWizard to generalize the classification on the EEG signal, irregardless of walking speed.

Figure 12 illustrates the setup as the participant walked on the treadmill while gazing at the SSVEP stimulus. Figure 13 summarizes the procedure of the experiment.

Corrupted channels/trials are automatically filtered (as described earlier under software pre-processing); The channel (1.5%) and trial rejection rate (2.8%) for WalkingWizard were comparable, or lower than all prior dry-electrode EEG headsets with published noisy channels / trials rejection rate [22, 48, 50]. The remaining 1851 trials, free from corrupted channels, were used to derive the subsequent results.

5.4 Results

5.4.1 Sample EEG and Noise Electrode Recording.

A sample of the recorded EEG and noise data (after preprocessing) is provided in Figure 14. In the sample EEG recording, it is noted that all five channels exhibit similar large amplitude fluctuations due to motion. The EEG channels contained high frequency signal (electrocortical signal) modulated on the noisy signal induced by motion, a signal that was observed to be absent from the noise electrode recording. The red box in Figure 14 highlights the observation described.

A spectrum plot of the same set of signals is provided in Figures 15(A) and 15(B). For each run of the experiment, there were 20 trials, of which there were 5 trials for each SSVEP stimulus, denoted as class number in each figure.

In Figure 15(A), the top spectrum plot showed the mean amplitude spectrum derived from the five trials of class 0 of the experiment (SSVEP frequency of 6.00 Hz). The bottom spectrum plot showed the same information as the top plot, after subtracting the mean amplitude spectrum of the noise electrode. Both plots contained coloured vertical lines indicating the frequency and harmonics of the four SSVEP stimuli. Figure 15(B) shows the corresponding spectral plot for class 1 of the experiment.

The top spectrum plot of Figure 15(A) showed the presence of SSVEP frequency at 12 Hz (second harmonics of 6 Hz), indicating that EEG signal could be recorded from the scalp despite the motion. While Figure 15(A) also showed a high frequency content at 6 Hz, this signal can also be observed in the top spectrum plot of Figure 15(B). It is only after subtracting the mean amplitude spectrum of the noise electrode then can we see from the bottom spectrum plot of both Figures 15(A) and 15(B) that the 6 Hz signal is a noise artifact, not the EEG recorded signal. The bottom spectrum plot of both Figures 15(A) and 15(B) clearly shows the SSVEP frequencies at the second harmonics (12 and 13.3 Hz, respectively) of the SSVEP stimuli.

Similar observations can be seen in the other experimental data where despite the top spectrum plot being largely overwhelmed by the large amplitude motion artifacts. The second spectrum plot showed that after subtracting the mean amplitude spectrum of the EEG signal from that of the noise electrode, the SSVEP frequencies became prominent, indicating the effectiveness of the noise electrode to remove motion artifact from the EEG recording.

We would like to highlight that for this sample EEG recording, the noise electrode data was highly correlated to the EEG data, allowing us to perform a direct subtraction of the amplitude spectrum to recover the EEG signal. For the recorded EEG data from other participants, we would require a weighted subtraction—a function that would be learnt by our classification algorithm.

5.4.2 Classification Performance.

The EEGNet model performance was evaluated using a 8-fold cross-validation approach, where the training and validation set consisted of the EEG data at static, 1.5, 3.0, 4.5, and 5.0 km/h of the other seven participants. Cross-subject accuracy performance is obtained as no user specific data is provided for training. No data from participant selected speed of 2.0, 3.5, and 4.0 km/h were used for training to evaluate if WalkingWizard’s model can generalize across all speeds, even at speeds not specifically trained for. Table 5 presents the results.

Table 5.

Participant	0	1.5	3	4.5	5	2	3.5	4
	Assigned Speed (km/h)					Selected Speed (km/h)
1	0.95	0.90	0.90	0.48	0.58	0.95	0.75
2	0.85	0.80	0.55			0.75*
3	0.70	0.85	0.75	0.43	0.65		0.60	0.70
4	1.00	0.95	1.00	0.83	0.90		0.95	0.85
5	0.88	0.98	0.95	0.84	0.76	1.00*
6	1.00	1.00	1.00	0.95	1.00			0.98*
7	1.00	1.00	0.98	0.80	0.80	1.00	1.00
8	0.95	0.88	0.85	0.88	0.85	0.95	0.90
AVG	0.92	0.92	0.87	0.74	0.79	0.91	0.84	0.88
S.D	0.097	0.069	0.146	0.189	0.133	0.106	0.146	0.115

Table 5. Cross-Subject Classification Accuracy of WalkingWizard

* For these entries, the participant selected the same speed twice for the experiment—shown results is the average of the two runs.

We compare the performance of WalkingWizard against other state-of-the-art dry-electrode EEG in literature in Figure 16. In Figure 16(A), we investigated the performance of WalkingWizard if we were to exclude the noise electrode, i.e., only use the four EEG electrodes’ data, for the training of EEGNet model. The results show that without the noise electrode, WalkingWizard exhibit similar performance as other state-of-the-art under slow motion, but pale in comparison at high walking speed.

Figure 16(C) provides a comparison of the cross-subject accuracy performance against other state-of-the-art.

We also compare the within-subject accuracy performance of WalkingWizard in Figure 16(B). Since each participant performed two runs of experiment at each of the five assigned speeds, we reserved one run of each assigned speed, plus both runs of the participant-selected speed, for the test set. The other run of each assigned speed was added to the initial train and validation set to train the model.

Figure 16 shows that WalkingWizard out-performed all other state-of-the-art dry-electrode EEG headset, achieving a high 87% within-subject classification accuracy at 5.0 km/h.

In Figure 16(D), we compare the classification performance of WalkingWizard between the Canonical Correlation Analysis (CCA) and the EEGNet algorithm.

Several observations can be noted from the results. First, WalkingWizard was able to classify the EEG signal to the SSVEP stimuli with accuracy of at least 74% for all walking speeds using EEGNet. Second, despite training EEGNet based on the assigned speed of 0, 1.5, 3.0, 4.5 and 5.0 km/h, the model was also able to accurately classify the EEG signal at other speeds (2.0, 3.5, and 4.0 km/h) that it was not trained with. This showed that the trained model was able to generalize to other walking speeds, implying that WalkingWizard could be deployed for natural walking motion. Finally, the results from the CCA algorithm showed that while it could potentially outperform EEGNet when the participants were not in motion, its performance was consistently inferior to that from EEGNet once the participants started walking. This could be attributed to the fact that important noise data captured by the noise electrode were not used in the CCA algorithm.

Table 5 and Figure 16 show that WalkingWizard is currently the best performing dry-electrode EEG headset for SSVEP experiments under motion. WalkingWizard also exhibited better performance in a static scenario—this could be attributed to lower electrode-scalp impedance due to the use of spring-loaded pins and flexible PCB substrate.

6 Application Demonstration

WalkingWizard was originally configured for SSVEP as it is the gold standard for comparing with other EEG headsets.

Here, to illustrate the potential of WalkingWizard as a truly wearable EEG, we demonstrated the versatility of WalkingWizard through several application demonstrations:

—

Perform SSVEP experiments in a VR environment

—

Monitor cognitive state in respond to environmental stimuli

—

Monitor cognitive state through a guided meditation exercise

6.1 Perform SSVEP Experiments in a Virtual Reality (VR) Environment

Given the growing significance and relevance of immersive technologies in today’s digital landscape, we explored the application of WalkingWizard in a VR environment. Moreover, extending from WalkingWizard’s distinctive advantage during movement, we wanted to examine the performance of WalkingWizard during walking scenarios within VR.

For this demonstration, we employed the Oculus Quest 2, allowing us to evaluate the synergy between WalkingWizard and a standard, commercially available VR platform. The integration aimed at ensuring that users can experience seamless VR interactions while concurrently capturing high-fidelity EEG data, even in walking scenarios. Our goal was to underline the potential of WalkingWizard as a versatile tool, ready for both research and entertainment purposes in the VR domain.

6.1.1 Development of WalkingWizard Integrated.

After we validated the performance of WalkingWizard versions 1 and 2 (Figure 6) in SSVEP experiments, we developed WalkingWizard Integrated (see Figure 17). WalkingWizard Integrated—designed using Reference [4] as reference, integrates the ADCs, compute module and wireless transceivers within our flexible PCB Assembly. It is in a truly wearable form-factor that can be easily integrated to another headgear for use in daily life. The critical components on WalkingWizard Integrated are provided in Table 6.

Table 6.

Critical Component	Component Part Number
Analog-to-Digital Convertor	ADS1299
Compute Module	PIC32MX250B
RF Transceiver	RFD22301
Electrode Pin	Mil-Max 0871-0-57-20-82-14-11-0

Table 6. Critical Components Used in WalkingWizard Integrated

WalkingWizard Integrated was built around the PIC32MX250B microcontroller and ADS1299 ADC to ensure accurate, low-noise and high-speed analog-to-digital conversion of EEG signals. The four combed dry electrode EEG channels input and noise electrode channel input are connected to the first five input channels of the ADC. We also incorporated a 10-pin PCB header connector, to accommodate an additional three EEG channel inputs. This connector also serves as the reference and bias inputs for ADS1299, and also provide additional signals for debugging purposes.

To streamline the firmware development process and updates for the PIC32MX250B microcontroller and RFD22301 RF transmission module, we integrated another header on the left side of the PCB Assembly to provide easy access to program the respective firmware to facilitate iterative development and testing.

For future integration requirements and scalability, we also integrated two more connectors on the right side of the PCB Assembly, dedicated as SPI interface.

WalkingWizard Integrated is powered using a 3.7 V Lithium Polymer battery, selected for its lightweight and durability, crucial for extended use scenarios.

6.1.2 VR Software Design.

In our experiment, we developed a virtual environment using Unity (version 2022.3.10f1) specifically tailored for the Oculus Quest headset. The scene was designed on the Universal Render Pipeline (URP), utilizing the standard URP scene settings to ensure optimal rendering performance and visual fidelity. We incorporated Unity’s built-in sample assets—WorkshopSet for our VR scene. Within this set, several items were augmented to serve as flickering stimuli: the PaintSupplies assets, namely, the 5G Bucket, flickered at a frequency of 7.5 Hz, and the 1G Bucket at 6.67 Hz. Additionally, Box 1 and Box 2 from the sample assets were set to flicker at frequencies of 6 and 8.57 Hz, respectively.

To achieve the desired flickering effect, a custom FlickeringMaterial was applied to these objects. This material was governed by a specialized Flicker Stimulus Script written in C#, which toggled the assets between black and white colors at the specified frequencies.

The scene in the VR is provided in Figure 18.

6.1.3 WalkingWizard Integrated Software Design.

BrainFlow [15] python library was used to initialize and get real-time EEG data from WalkingWizard Integrated. It is also used in our implementation of digital filtering to pre-process the EEG recording.

Prior to the using WalkingWizard Integrated in the VR environment, we needed to fine-tune the EEGNet model trained in Section 5. First, we requested the user to put on WalkingWizard Integrated and the VR headset. Then, we requested the user to follow the experiment protocol in Figure 13 to perform two runs of the SSVEP experiment while walking on the treadmill at 3.0 km/h—with two minor deviations: (a) that the stimulus is continuously generated in the VR headset, and (b) that the user is prompted by a experimenter to switch gaze to one of the four objects once every 7 s.

WalkingWizard Integrated was built to replace the analog bandpass filter and notch filter with its digital equivalent. We performed digital bandpass filter of 5 to 30 Hz using a fourth order Butterworth filter without windowing. Then, we removed 50 Hz environmental noise using bandstop filter from 48 to 52 Hz, also with a fourth-order Butterworth filter without windowing. After the filtering is performed, we followed the same pre-processing steps in Section 3.7.1 to pre-process the EEG recording. It is to be noted that the same pre-processing steps can be performed here as the sampling rate of WalkingWizard Integrated is at 250 Hz, the same as that for gNAUTILUS Multipurpose.

Following the same data augmentation approach as described in Section 3.7.2, the pre-processed EEG recording is then used to fine-tune the trained EEGNet model in 5. This process was performed offline.

The fine-tuned EEGNet model was then used for inference in real-time.

6.1.4 Demonstration and Results.

WalkingWizard Integrated is put on under a commercial headband to provide mechanical support. Oculus Quest 2 is donned over the headband for the demonstration. Figure 19(A) illustrates the demonstration setup.

In the demonstration, the user randomly selects one of the four flickering objects to fix his gaze at. After approximately 3 s, the user announces the object number (left most object indexed as zero) that he gazed at. This object number is captured as the true output. The experimenter checks the inference output and capture it as the predicted output. The process was repeated for approximately 20 predictions. Fluctuations in predicted output before the user announces the object number were ignored.

The same demonstration was performed for another user, also after fine-tuning the EEGNet model.

The accuracy performance for the two users are 86% and 88%, respectively. The confusion matrices for both users are provided in Figure 19(B).

6.2 Monitor Cognitive State in Respond to Environmental Stimuli

We reconfigured WalkingWizard for real-time sensing of the cognitive state of a user as he goes about his everyday activities.

Specifically, WalkingWizard was reconfigured to monitor FAA—widely cited in literature to measure the approach-withdrawal attitude of two users to stimuli [27, 31, 57]—while walking and cycling around the neighbourhood. Employed broadly in neuromarketing research, FAA was used to study the effects that various factors (e.g., packaging, color, shape) could have on the emotions on the consumers; and in this demonstration the emotion of the user in different real-world settings.

The FAA index provides an indication of the hemispheric asymmetry in the alpha band (8–12 Hz) of the frontal cortex, and is calculated based on the formula in Reference [72]. The FAA index is reflected in real-time on the tablet (Microsoft Surface Go 3) using colour indication.

6.2.1 Hardware Reconfiguration.

For the first user, two of WalkingWizard’s flexible PCB assemblies were used to monitor the EEG of the user for this demonstration. One of the flexible PCB assembly monitored the F4, F8, and AF8 electrode position on the right hemisphere, while the other flexible PCB assembly monitored the F3, F7, and AF7 electrode position on the left hemisphere. The noise electrode for both flexible PCBs were used for reduction of motion artifacts. The ground and reference electrodes were both positioned at the FPz electrode position using the single electrode flexible PCB assembly. The wireless transceiver from g.NAUTILUS Multipurpose was hidden under the tongue of the baseball cap—Figure 21 shows the user wearing a cap with a thick tongue, where the wireless module is hidden under. The top and underside of WalkingWizard is shown in Figure 20.

As we only have one set of g.NAUTILUS Multipurpose, WalkingWizard Integrated was used for the second user. WalkingWizard Integrated was placed on the forehead of the user, with the original O1 and O2 electrodes now used to monitor the FP2 and FP1 electrode position, respectively. The ground and reference electrode were positioned at the left mastoid, but re-referenced to the FPz electrode position.

6.2.2 Software Design for Demonstration.

Two threads run concurrently on the Microsoft Surface Go 3 for the demonstration. The first thread records the EEG data from WalkingWizard at 15 s intervals on the non-volatile memory on the device. The second thread polls (10 s interval) the allocated directory for new files created by the first thread. With each new file created, the second thread: (1) performs preprocessing to remove corrupted EEG recording (based on amplitude threshold and correlation of channels—see Section 3.7), (2) performs Fast Fourier Transform (FFT) to get the power spectrum of the EEG recording, filtered to the alpha band (8–12 Hz), corrected with the spectral from the noise electrode, (3) calculates the FAA index, and (4) updates the colour display on the tablet—implemented using tkinter library, to indicate the cognitive state of the user, spanning from green, which indicates high FAA index (approach attitude), to red, which indicates low FAA index (withdrawal attitude).

6.2.3 Demonstration.

The demonstration trial was conducted on 15 October 2023 at 09:30 a.m. The users travelled a total distance of 8.1 km (5.0 km on bicycle, 3.1 km on foot) in around 88 min with WalkingWizard / WalkingWizard Integrated, Surface Go 3 tablet, a bicycle and a GoPro (to record footage of demonstration) attached to the backpack (see Figure 21). The breakdown of the journey is provided in Table 7.

Table 7.

6.2.4 Results.

The FAA index of the users is provided in Figure 22. A yellow background indicates journey on bicycle, while that journey on foot is indicated with a white background. A high value of FAA indicates a positive (approach) emotion due to higher activation (indicated by lower alpha band) on the left hemisphere, while a low value of FAA indicates a negative (withdraw) emotion.

The users were surveyed after the experiment to gather insights on their emotional state through the experiment. We matched their response against the FAA trend that were captured. (Interval 1) FAA index dropped rapidly as the users were feeling apprehensive about the demonstration. (Interval 4 to 5) FAA index improved significantly as the users felt accomplished that WalkingWizard was working smoothly. (Interval 6) FAA index dropped drastically as the pathway was extremely bumpy, making the users worried that EEG data collection would be heavily corrupted by motion artifacts. (Interval 7) FAA index improved significantly while walking in a shaded serene park beside a reservoir. WalkingWizard maintained real-time sampling and sensing of the users’ cognitive state throughout the journey, attesting to its robustness.

The FAA index plot in Figure 22 shows the potential for WalkingWizard and/or WalkingWizard Integrated in monitoring the cognitive state of the users while on the move. Our findings indicated a positive correlation between exposure to green space and FAA index [8, 10] and negative correlation between anxiety and FAA index [17, 23].

The difference in magnitude in the the FAA index, especially from Intervals 5 onwards, could either be attributed to the difference in electrode positions used for monitoring, or to the different hardware used. This would be investigated in further details in future.

6.3 Monitor Cognitive State Through a Guided Meditation Exercise

The efficacy of WalkingWizard Integrated was demonstrated in an experiment aimed at monitoring the cognitive states of the user while listening to a radio broadcast (control situation) and engaging in a guided meditation session.

Emulating the experiment in Reference [9], the user was exposed to a standard radio broadcast (BBC World Service) for a duration of 12 min, seated with both eyes closed. The aim was to emulate an everyday passive listening scenario, where the cognitive state remains relatively constant and is not subjected to intentional modulation.

For the guided meditation session, the user (not trained in meditation) was instructed to follow the meditation protocol from a popular meditation podcast—“Guided Meditation: Deep Relaxation” by “The Honest Guys” for a duration of 12 min. The user is seated with both eyes closed. Guided meditation, by design, is targeted to alter the cognitive state of the individual, leading to deeper states of relaxation, focus, and mindfulness.

6.3.1 Software Design for Demonstration.

EEG recording from WalkingWizard Integrated was retrieved using the Brainflow [15] python library. Following the procedure of Reference [9], we performed digital bandpass filter of 2 to 30 Hz and calculated the band power of the respective bands—Theta (4–8 Hz), Alpha (8–12 Hz), and Beta (12–30 Hz).

6.3.2 Results.

The spectrogram plot derived from the O1, Oz, and O2 electrodes of WalkingWizard Integrated for both the guided meditation and control experiment is provided in Figure 23.

The spectrogram plot allows us to monitor how the different frequency bands of the user changed through time. Since the user’s eyes were closed, high alpha wave can be seen in both spectrogram—denoted by the bright yellow band from around 10–12 Hz. In the guided meditation, the band power of the alpha wave was significantly higher, and for longer duration than in the control condition. This provides objective evidence that guided meditation improve relaxation for the user.

We compared the band-power of Theta (4–8 Hz), Alpha (8–12 Hz), and Beta (12–30 Hz) for Guided Meditation against Control—listening to radio broadcast in Figure 24. While we were unable to compare the difference in band-power across all regions of the head, our results on the occipital region of the head align closely with other literature [9, 34, 36, 38, 54, 66] that reported increase in the Theta to Beta band-power during meditation.

7 Discussion and Future Work

Software optimizations. In WalkingWizard, we focused on the effect that novel hardware has on the performance of a dry-electrode EEG headset—using simple thresholds and correlation for the software pre-processing steps. In literature, the ASR algorithm [58] was cited in numerous publications [20, 22, 55, 58, 59] as a software pre-processing step to clean the EEG recording prior to subsequent processing. In Reference [63], Ferris et al. explored the inclusion of ASR algorithm as pre-processing step to their dual-electrode configuration on high density wet-electrode EEG. As the ASR algorithm could be used in real-time, a logical extension of WalkingWizard would be the inclusion of ASR algorithm as a pre-processing step (replacing the gross artifact rejection method) to clean the EEG signal. This may further improve WalkingWizard’s accuracy.

The software compute platform of WalkingWizard can clearly be substantially optimized and miniaturized further, with off-the-shelf or custom-designed system-on-chips [37] to perform computation directly [12, 74] on the EEG cap. This will also obviate the need for the wireless back-end and reduce overheads.

Lifetime Performance. As WalkingWizard was designed for everyday use and built on flexible PCB, we wanted to ascertain if the performance of WalkingWizard will degrade over time. In Section 4.2, the two users recruited for the alpha rhythm test performed the experiment 4 months away from each other. From the spectrograms provided in the Appendix A.2, it can be observed that there is no significant difference between the two sets of 1 min recording. We infer that there is no significant degradation in the performance of WalkingWizard in the 4 months interval.

Mobile EEG applications. We believe WalkingWizard’s ability to continuously sense our cognitive state as we go about our everyday life will enable transformative applications. WalkingWizard provides a quantitative measure of our cognitive state (stress/relax, likes/dislikes, emotions), which can be correlated with our environment, objects or interactions with others. Doctors could use EEG readings to assess the impact of behavioral interventions for wellness such as “park prescriptions,” marketers could use them to fine-tune the advertisements, and dating agencies/app could use them to improve their match.

WalkingWizard also unlocks a new modality for us to convey our intent—to a device, or to another person. Cyclists could switch the music track without getting their hands off the handle, paramedics could update on the status of the patient while still attending to him, and tactical forces could stealthily synchronize their assault without any observable action.

Clinical Applications. Existing EEG systems used in clinical settings take a long time to setup, and can only be performed by expert technicians. WalkingWizard—with its accurate sensing capabilities using dry EEG electrodes, in face of practical motion artifacts, exhibits potential to overcome the huge healthcare gap and enables continuous EEG monitoring of patients at home in a natural form-factor, lending to wide range of potential healthcare applications in physical and mental health.

8 Conclusion

We presented a comprehensive literature survey on the current state-of-the-art EEG headsets and identified the research gap of a truly wearable EEG headset for everyday use. We presented the design considerations in our proposed design, fabrication and prototyping of dry dual-electrodes on flexible PCBs mounted atop a baseball cap. We then described the full system hardware-software implementation of WalkingWizard. Hardware performance testing—electrode-scalp impedance and ability to capture EEG alpha rhythm, were performed and results compared against both wet EEG, and commercially available dry EEG headsets. In our evaluation of WalkingWizard, 8 participants were recruited to perform SSVEP experiments, achieving a high classification accuracy of 87% for walking speeds up to 5.0 km/h. We also presented our work on developing WalkingWizard Integrated, and used it to demonstrate several applications as proof-of-concept: (1) To classify SSVEP stimuli in a VR environment while walking, (2) To record and display the mental state of the user (FAA) in real-time while moving around the neighbourhood, and (3) To compare the effect of guided meditation against a control condition. We envision WalkingWizard enabling continuous sensing of our intent and wellbeing in our everyday lives.

Acknowledgments

The authors express our sincere gratitude to Ananta Narayanan Balaji for his invaluable assistance in reviewing this article.

A Hardware Performance Spectrogram

A.1 Comparison between WalkingWizard and Wet Electrode

The spectrogram comparing the performance of WalkingWizard and wet electrode is provided in Figure 25 below.

A.2 Comparison between WalkingWizard and Other Dry EEG Electrodes

The spectrogram to evaluate the performance of WalkingWizard against other dry EEG headsets in capturing alpha rhythm while staying static is provided in Figures 26 and 27. The 1 min recording from each user is concatenated to generate the spectrogram of 2 min duration.

Fig. 26.

Fig. 27.

The same set of spectrogram, while the users are in a vibrating vehicle static is provided in Figures 28 and 29. The 1 min recording from each user is concatenated to generate the spectrogram of 2 min duration.

Fig. 28.

Fig. 29.

References

[1]

Guideline twelve: guidelines for long-term monitoring for epilepsy. 2008. Journal of CLinical Neurophysiology 25, 3 (June 2008).

Abstract

1 Introduction

2 Related Works

3 System Design and Implementation

3.1 Electrode Pin

3.2 Number of Electrodes

3.3 Noise Electrode

3.4 Dry Dual-Electrode Flexible PCB Assembly

3.5 Analogue Filters, ADC, Wireless Transceivers

3.6 WalkingWizard Overview

3.7 Software Design

3.7.1 Pre-processing.

3.7.2 Classification.

4 Hardware Performance

4.1 Performance Against Gold-standard Wet EEG Electrode

4.1.1 Impedance.

4.1.2 Alpha Rhythm Capture.

4.2 Performance Against Other Dry EEG Headset

4.2.1 Impedance.

4.2.2 Alpha Rhythm Capture.

5 Evaluation

5.1 Participants

5.2 EEG Data Acquisition

5.3 Experiment Protocol and Procedure

5.4 Results

5.4.1 Sample EEG and Noise Electrode Recording.

5.4.2 Classification Performance.

6 Application Demonstration

6.1 Perform SSVEP Experiments in a Virtual Reality (VR) Environment

6.1.1 Development of WalkingWizard Integrated.

6.1.2 VR Software Design.

6.1.3 WalkingWizard Integrated Software Design.

6.1.4 Demonstration and Results.

6.2 Monitor Cognitive State in Respond to Environmental Stimuli

6.2.1 Hardware Reconfiguration.

6.2.2 Software Design for Demonstration.

6.2.3 Demonstration.

6.2.4 Results.

6.3 Monitor Cognitive State Through a Guided Meditation Exercise

6.3.1 Software Design for Demonstration.

6.3.2 Results.

7 Discussion and Future Work

8 Conclusion

Acknowledgments

A Hardware Performance Spectrogram

A.1 Comparison between WalkingWizard and Wet Electrode

A.2 Comparison between WalkingWizard and Other Dry EEG Electrodes

References

Cited By

Index Terms

Recommendations

Designing a Sequential Hybrid BCI Speller on Fusion of Motor Imagery and Steady-State Visually Evoked Potentials EEG Data

Feature Analysis of EEG Based Brain-Computer Interfaces to Detect Motor Imagery

Comprehensive review on brain-controlled mobile robots and robotic arms based on electroencephalography signals

Comments

Information

Published In

Publisher

Publication History

Check for updates

Author Tags

Qualifiers

Funding Sources

Contributors

Other Metrics

Bibliometrics

Article Metrics

Other Metrics

Citations

Cited By

View options

PDF

eReader

Login options

Full Access

Figures

Other

Share

Share this Publication link

Share on social media