EEGMobility: Mobile EEG Systems
- eegMobility is a systems framework that integrates mobile EEG recording with peripheral biosignals to capture neural dynamics during locomotion and navigation.
- It leverages various architectures—from all-in-one workstations like the PiEEG Box to distributed smartphone stacks—for real-time artifact mitigation and synchronized signal acquisition.
- The approach advances assistive control, gait decoding, and vehicular interfaces by emphasizing lightweight on-device inference and adaptive calibration under real-world conditions.
Searching arXiv for relevant papers on mobile EEG and eegMobility. eegMobility denotes mobile and ambulatory electroencephalography systems that record, synchronize, and interpret neural activity during locomotion, navigation, or control of mobility-related devices. In the available literature, the term is used both for the PiEEG Kit’s mobile EEG platform and for the real-time fusion of ambulatory EEG with peripheral biosignals in mobility assistance. The associated work spans all-in-one Raspberry Pi workstations, smartphone-based neuroimaging, treadmill and field protocols, wheelchair and driving interfaces, and synchronized brain–body sensing for gait and sport. This suggests that eegMobility is best understood as a systems framework rather than a single device class (Rakhmatulin, 5 Mar 2025, Saitis et al., 2018).
1. Architectural forms and hardware organization
Two recurrent architectural forms appear in eegMobility research. One is the integrated mobile workstation, exemplified by the PiEEG Box, described as an all-in-one mobile EEG workstation. Its major hardware building blocks are a Raspberry Pi (Model 4B) single-board computer, a PiEEG Shield with the TI ADS1299 24-bit analog front end for up to eight simultaneous EEG channels plus auxiliary biopotential inputs, an electrode-cap interface, a 3.5″ TFT display via SPI for real-time visualization and control, an I²C-attached sensor module, a 5 V Li-ion power module, and standard peripherals. The enclosure footprint is 27 cm × 17.5 cm × 6.5 cm, and the weight is ~0.8 kg (Kit w/o cap), with shoulder-strap slots, a Velcro panel, ventilation slots, and a front-panel weather-resistant seal around the display window (Rakhmatulin, 5 Mar 2025).
A second form is the distributed mobile stack, in which sensing, computation, and feedback are separated across commodity devices. The Smartphone Brain Scanner combined an Emotiv EPOC headset with smartphone or tablet hardware and a Qt 4/5 in C++ framework, and was presented as the first fully mobile system for real-time 3D EEG imaging. Its software stack separated data acquisition, data processing, and application layers, and included digital filtering, FFT, CSP, LDA, and Bayesian Minimum-Norm and LORETA source reconstruction (Stopczynski et al., 2013).
A more recent pocketable configuration used two identical Android smartphones, a Smarting Pro wireless EEG system, a Movella Dot wrist sensor, and Lab Streaming Layer to capture EEG, video, pose, and inertial data during basketball free-throw shooting. One tripod-mounted phone ran the Smarting Pro app for EEG + video, while the second ran a MediaPipe Pose Landmark Detection app; all streams were tagged and merged in a single XDF file via LSL RECORDA (Contreras-Altamirano et al., 9 Jan 2025).
| System | Core acquisition stack | Mobility context |
|---|---|---|
| PiEEG Box | Raspberry Pi 4B, PiEEG Shield, ADS1299, 8-channel dry cap, on-board display | all-in-one mobile EEG workstation |
| Smartphone Brain Scanner | Emotiv EPOC, smartphone/tablet, Qt C++ framework | real-time 3D EEG imaging |
| Two-smartphone portable setup | Smarting Pro, two Android phones, MediaPipe Pose, LSL | simultaneous human movement and mobile EEG acquisition |
2. Acquisition modalities, synchronization, and quantitative instrumentation
eegMobility systems differ substantially in electrode count, sensor topology, and acquisition bandwidth, but they share an emphasis on synchronized brain–body measurement. In the PiEEG Box, the cap is an 8-channel dry Ag/AgCl electrode cap with 10–20 system positions and a configurable montage using single-ended Fz, Cz, Pz, and Oz plus mastoid reference and common bias. The ADS1299 front end provides 24-bit resolution, input noise ~1 µVpp, programmable gain from 1× to 12×, 250 SPS default sampling up to 2 kSPS, input dynamic range of ±4.5 mV with gain=12, and typical CMRR >110 dB. The report defines the principal instrumentation metrics as and Electrode–skin impedance is typically <50 kΩ for dry Ag/AgCl at initial setup, and is monitored via test drive in the GUI. Wireless connectivity includes Wi-Fi 802.11ac and Bluetooth 5.0 LE, with TCP/IP sockets or WebSocket over HTTP(S) for streaming; at 250 SPS × 8 channels × 3 bytes/sample, the data rate is ≈ 6 kB/s, with ~20–50 ms round-trip latency under typical lab Wi-Fi conditions (Rakhmatulin, 5 Mar 2025).
Mobile EEG datasets often add dense peripheral instrumentation. One scalp/ear-EEG dataset recorded 32-channel scalp-EEG, 14-channel ear-EEG, 4-channel electrooculography, and three 9-channel inertial measurement units during standing, slow walking, fast walking, and slight running. EEG and EOG were sampled at 500 Hz and IMUs at 128 Hz, with all data down-sampled to 100 Hz in preprocessing; the dataset follows EEG-BIDS and BrainVision Core Data Format (Lee et al., 2021).
In sports-related eegMobility, synchronized alignment across heterogeneous time bases is central. In the portable free-throw setup, EEG timestamps were treated as the reference timeline, and pose and IMU streams were upsampled to 250 Hz by linear interpolation:
The MATLAB function interp1 executed this dejittered alignment (Contreras-Altamirano et al., 9 Jan 2025).
Vehicular eegMobility extends the same principle to larger sensor suites. In real-world on-road driving, a Renault Twizy was instrumented with wireless 16-channel OpenBCI Cyton + Daisy EEG at 125 Hz, GNSS/INS at 100 Hz, LiDAR + IMU at 10 Hz, and stereo cameras at 30 Hz. Synchronization was performed via ROS Noetic, with EEG serving as the reference timeline and vehicle kinematics linearly interpolated to EEG sample times for exact pairing of neural and motion data (Alosaimi et al., 21 Apr 2026).
3. Signal conditioning and motion-artifact management
Motion artifact is a defining systems problem in eegMobility. Hardware-level mitigation in the PiEEG Box includes differential acquisition via ADS1299 to eliminate common-mode cable noise, a driven-right-leg electrode to reduce body-EM interference, and an on-board RC low-pass with cutoff ~500 Hz to suppress high-frequency cable microphonics. The accompanying SDK includes real-time band-pass filtering, ICA, wavelet decomposition, EMD, and CCA, with an example filter chain of Butterworth band-pass (0.5–45 Hz), notch filter at 50/60 Hz, and ICA plus component rejection. In a continuous 2-hour ambulatory treadmill test at 5 km/h, motion artifacts were reported as <15 µVpp after DRL plus software filters, and post-processing residual RMS fell from RMS_motion ≈ 20 µV to RMS_residual ≈ 5 µV, corresponding to a 75% reduction with the filter chain (Rakhmatulin, 5 Mar 2025).
Standardized mobile BCI preprocessing pipelines often rely on conservative filtering and automated channel repair. In the scalp/ear-EEG locomotion dataset, preprocessing used a 5th-order Butterworth high-pass >0.5 Hz, automated line-noise removal, EOG regression via BCILAB’s flt_eog, bad-channel detection with flt_clean_channels, spherical spline interpolation, and common-average re-referencing. On average 2.4 scalp channels and 1.4 ear channels were interpolated per session. Cluster-based permutation testing showed significant increases in low-frequency delta power across most channels as speed increased, reflecting gait artifacts (Lee et al., 2021).
A recurrent controversy concerns whether increasingly aggressive artifact removal improves decoding. In on-road driving, eleven preprocessing strategies were compared, ranging from baseline z-score normalisation only to combinations of PyPREP, RANSAC, ICA, ASR, and AutoReject. The minimal pipeline, defined as z-score normalisation only, yielded the highest Macro-F1 and balanced accuracy, whereas more aggressive pipelines consistently degraded performance. The authors interpreted this as evidence that real-world EEG carries low-SNR yet discriminative intention-related activity that can be lost when over-cleaned (Alosaimi et al., 21 Apr 2026).
Mobile cognitive classification during walking shows a complementary strategy: heavy offline cleaning followed by architecture-level robustness. In treadmill oddball experiments, preprocessing used band-pass filtering, AMICA, dipole fitting, removal of components dominated by eye and muscle artifacts, average re-referencing, and baseline correction, while the adapted CN-EEGNet replaced ELU with Mish, removed one spatial dropout layer, and added a second separable-convolution layer. Across seated, unloaded walking, and loaded walking conditions, classification remained high and stable (Cichy et al., 2023).
4. Mobility paradigms and neural signatures
A major strand of eegMobility research studies the spectral and event-related signatures of movement, stopping, and locomotor pathology. In Parkinson’s disease, ambulatory EEG during Timed Up-and-Go tasks was analyzed with event-related spectral perturbation, defined as
Transitions to freezing of gait showed significant elevations in alpha and part of beta power at Cz and O1 versus normal walking, whereas voluntary stopping exhibited a global surge across delta, theta, alpha, and beta during the transition period. Critically, during the 0→+2 s stopping window, voluntary stopping showed significantly lower delta and low-beta power at Cz than freezing of gait, while freezing maintained elevated delta and low-beta. The paper presents this dissociation as a spectral fingerprint for discriminating involuntary freezing from purposeful stopping (Cao et al., 2021).
Speed-dependent degradation of canonical BCI responses is well documented but not uniform. In the scalp/ear-EEG locomotion dataset, scalp ERP target-vs-non-target AUC fell from 0.90 ± 0.07 at 0 m/s to 0.67 ± 0.07 at 1.6 m/s, and scalp SSVEP accuracy fell from 88.7 ± 19.5 % at 0 m/s to 80.7 ± 20.4 % at 1.6 m/s. Ear-EEG performance was lower, but remained viable. The same study reported significant speed-dependent spectral shifts in cortical regions relevant to ERP and SSVEP tasks, interpreted as evidence of dual-task interference (Lee et al., 2021).
At the same time, mobile EEG does not reduce exclusively to degraded laboratory paradigms. In treadmill oddball experiments, the adapted CN-EEGNet achieved mean accuracies of 99.43 % in seated unloaded, 96.09 % in walking unloaded, 96.17 % in walking loaded, and 99.04 % in seated loaded conditions, with overall mean ≈ 97.2 %. These results were presented as the first documented implementation of a deep neural network for classifying cognitive neural state during dual-task walking (Cichy et al., 2023).
Portable brain–body imaging has also captured preparatory activity preceding self-paced movement. In basketball free-throw shooting, grand-average readiness potential was evident over fronto-central channels, with significant bins at Cz from −400 ms to −100 ms, and maximum negativity at Fz. No significant point-biserial correlation was found between readiness-potential amplitude and shot outcome, but pose-derived landmarks revealed significant upper-body differences for ten participants (Contreras-Altamirano et al., 9 Jan 2025).
eegMobility can also target affective and environmental load rather than explicit motor commands. In visually impaired mobility, multimodal classification using EEG and peripheral biosignals reached outdoor multi-class wAUROC = 0.93 and indoor multi-class wAUROC = 0.87, with mean tonic EDA, heart rate, and EEG delta ERD/ERS at F3, F4, T7, T8, P7, and P8 among the top predictive biomarkers (Saitis et al., 2018).
5. Assistive control, gait decoding, and mobility interfaces
Wheelchair control is one of the most direct assistive embodiments of eegMobility. One non-invasive BCI system used an Emotiv EPOC+ headset and the Emotiv Cortex SDK to learn user-specific patterns for four discrete commands—Push, Pull, Left turn, and Right turn—and streamed classified commands to a Unity 3D wheelchair simulation via WebSocket. Calibration comprised a neutral baseline and four mental tasks with 3D “cube control” feedback until classification accuracy exceeded 80%, with total training time of ≈15–20 minutes per user. In a pilot group of 5 able-bodied participants and 2 participants with limited upper-limb mobility, task completion time was 45 s for EEG control versus 40 s for joystick, and NASA-TLX workload was 55/100 versus 40/100 (Ghasemi et al., 2024).
A lower-cost implementation used a Neurosky MindWave Mobile headset, Android smartphone, HC-06 module, Arduino UNO, ultrasonic sensor, MQ-2 smoke sensor, and a seat push-button. Control logic relied on “attention” and “blink strength,” with single blink initiating a direction cycle, double blink within 400 ms confirming the current direction, and response time of ≈ 0.4 s from double blink to wheelchair motion. In three indoor trials each, collision avoidance, smoke alerts, and fall alerts each reached 100%, and the combined control plus safety system showed zero collisions in 10 obstacle runs (Sarkar et al., 6 Jan 2025).
Another wheelchair prototype used an Emotiv EPOC+ headset, LattePanda, DWT with db4 mother wavelet, SVM classification, and Arduino Mega motor control. The dataset comprised 5 commands × 50 trials per command × 8 s per trial, and the SVM achieved 70% overall accuracy on a held-out test set, with 68% overall accuracy in online evaluation. Ultrasonic sensors on the front and back overrode commands to “Stop” when an obstacle was detected closer than a threshold (AlAbboudi et al., 2020).
Beyond discrete assistive commands, eegMobility increasingly targets continuous movement decoding. EEG2GAIT processed 59 EEG channels with a Local Temporal Learner, a Graph Construction Module, a Hierarchical GCN Pyramid, a Global Spatial Learner, and a Global Temporal Learner, and optimized joint-angle prediction with a Hybrid Temporal–Spectral Reward loss. On the new Gait-EEG Dataset of 50 participants, EEG2GAIT reached and ; on the MoBI dataset, it reached and . Saliency maps localized the most influential electrodes to Cz, FC1/2, Fz, and CP1/2 (Fu et al., 2 Apr 2025).
Driver-intention decoding extends the same logic to vehicular control. In real-world on-road driving, twelve deep-learning architectures were evaluated, and TSCeption achieved the highest average accuracy of 0.907 and Macro-F1 of 0.901, with <1.5% drop at Δ = 1000 ms relative to baseline. Peak performance occurred in the 400–600 ms interval, interpreted as a critical neural preparatory phase preceding driving manoeuvres (Alosaimi et al., 21 Apr 2026).
Portable competitive BCI further broadens the domain. A modular online system built for the Cybathlon used mBrainTrain Smarting mobi, Lab Streaming Layer, ZeroMQ/UDP, CSP plus Morlet wavelet features, and three diagonalized structured state-space sequence layers. It used three mental and motor imagery classes to control up to five control signals, achieved 84.0% offline accuracy for one pilot, and 73.2% success rate in real-time gameplay after the competition (Tscherniak et al., 28 Nov 2025).
6. Constraints, deployment trade-offs, and emerging directions
The principal constraints of eegMobility are mechanical, computational, and human factors rather than acquisition alone. In the PiEEG Box, the reported limitations were dry-electrode impedance drift during long motion tasks, limited onboard storage and CPU headroom for heavy real-time artifact removal, and cable sway artifacts due to separate cap cabling. Proposed enhancements included an IMU tri-axial accelerometer/gyroscope on the EEG cap for motion-adaptive filtering, active-shielded electrode cables, enclosure miniaturization into a flexible belt-pack form factor, C++/NEON firmware acceleration for on-device ICA, and ultra-low-power Raspberry Pi alternatives such as Compute Module 4 (Rakhmatulin, 5 Mar 2025).
Earlier mobile neuroimaging platforms reported a related but distinct constraint set. The Smartphone Brain Scanner measured end-to-end latency at 80–125 ms with jitter of 16–26 ms, noted that Android is a non–real-time OS, and highlighted drying of dry or saline electrodes over hours, lower channel density, and limited on-device removal of muscle and EOG artifacts. Proposed mitigations included online adaptation of noise and source hyperparameters, offloading ICA or CSP to a local laptop or cloud server, and trade-off sliders for visualization quality versus frame rate (Stopczynski et al., 2013).
Resource-efficient inference has therefore become central to eegMobility. Q-EEGNet showed that 8-bit fixed-point quantization of EEGNet incurred only a 0.4% accuracy loss on 4-class motor imagery while yielding 64× speedup and up to 85% reduction in memory footprint on the Mr.Wolf PULP SoC. At 350 MHz and 1.2 V, inference time was 5.82 ms, energy per inference was 0.627 mJ, and energy efficiency was 21.0 GMAC/s/W, reported as 256× more energy-efficient than an EEGNet implementation on an ARM Cortex-M7 (Schneider et al., 2020).
A related model-design perspective appears in EEGMobile, which replaced a large transformer backbone with MobileViTV2 and used knowledge distillation for EEG regression. On the EEGEyeNet Absolute Position Task, EEGMobile obtained RMSE 53.6 ± 0.6 mm, was 33% faster than EEGViT-TCNet, and 60% smaller in parameter count, while remaining only ~3 % worse in RMSE than the prior state of the art. The paper explicitly framed this as applicable to resource-constrained devices and mobile-friendly models for EEG regression (Liang et al., 2024).
Human-centred adaptation remains equally important. In the Cybathlon-oriented mobile BCI, reduced performance in competition was attributed primarily to stress and the challenging competition environment, and the authors recommended short training times <20 min, real-time latency <250 ms, adaptive calibration, user-driven task selection, and mobile/web feedback for at-home training (Tscherniak et al., 28 Nov 2025).
Taken together, these results suggest that eegMobility is moving toward multimodal synchronization, lightweight on-device inference, and adaptive calibration under real motion and environmental load, rather than toward a single canonical preprocessing stack or hardware configuration.