⚡ electrical engineering

WearBCI Dataset: Understanding and Benchmarking Real-World Wearable Brain-Computer Interfaces Signals

This paper introduces WearBCI, the first comprehensive dataset featuring synchronized EEG, IMU, and egocentric video recordings from 36 participants under diverse motion dynamics, designed to benchmark wearable BCI signal quality and evaluate motion artifact mitigation techniques for real-world applications.

Original authors: Haoxian Liu, Hengle Jiang, Lanxuan Hong, Xiaomin Ouyang

Published 2026-04-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Haoxian Liu, Hengle Jiang, Lanxuan Hong, Xiaomin Ouyang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

🧠 The Big Idea: Listening to the Brain While You Move

Imagine you have a super-sensitive microphone that can hear your brain's thoughts. This is what a Brain-Computer Interface (BCI) does. It translates brain waves into commands for computers, helping people control devices, recover from strokes, or monitor their health.

The Problem:
Until now, these "brain microphones" have been like high-end recording studios: they require the user to sit perfectly still in a quiet room with a messy tangle of wires and wet gel on their head. If you move, the recording gets ruined by static and noise. It's like trying to record a symphony while someone is shaking the microphone and stomping on the floor.

The Solution:
The researchers at the Hong Kong University of Science and Technology created WearBCI. Think of this as a "field recording" project. They built a portable, dry-electrode headset (like a cool headband) and asked 36 people to wear it while they did normal, messy human things: walking, nodding, typing, and navigating a room.

🎒 What They Collected (The "Three-Legged Stool")

To understand how movement messes up brain signals, they didn't just record the brain. They recorded three things at the exact same time, like a three-legged stool that keeps everything balanced:

The Brain (EEG): The actual thoughts and neural activity.
The Motion (IMU): Sensors on the head, wrists, and ankles that act like a "motion detective," tracking every shake, step, and turn.
The View (Video): A camera on the person's chest (like a GoPro) showing exactly what they were seeing and doing.

The Analogy:
Imagine trying to figure out why a song sounds distorted.

Old way: You only listen to the distorted song. You have no idea if the singer is off-key or if a truck just drove by.
WearBCI way: You listen to the song plus a video of the truck driving by plus a sensor measuring the vibration of the floor. Now you know exactly why the song is distorted.

🏃‍♂️ The Experiments: From "Sitting Still" to "Chaos"

The researchers tested the headsets in three levels of difficulty, like a video game with increasing levels:

Level 1: The Chill Zone (Body Movements): People sat still but did small things like shaking their heads, blinking, or stretching their arms.
- Result: Small movements caused small glitches, mostly in specific parts of the brain signal.
Level 2: The Jog (Walking): People walked at slow, medium, and fast speeds.
- Result: Walking created a rhythmic "thumping" noise in the brain signal, like a drumbeat that drowned out the music. The faster they walked, the louder the drumbeat.
Level 3: The Adventure (Navigation): People walked through a room, opened doors, greeted a person, and avoided obstacles.
- Result: This was the "boss level." Sudden stops, turns, and interactions created the most chaotic noise.

🔍 What They Discovered

1. Motion is the Enemy of Clarity
They found that as movement gets more intense, the "brain signal" gets covered in static. It's like trying to hear a whisper in a windstorm. The wind (motion) creates so much noise that the whisper (brain signal) gets lost.

2. The "Cleaners" Have Mixed Results
The team tested various software tools (algorithms) to try to "clean" the noisy brain signals, like using Photoshop to remove dust from a photo.

Old School Cleaners (Traditional Math): These worked okay when people were sitting still, but when people started moving, they either failed completely or accidentally deleted the actual brain thoughts along with the noise.
New School Cleaners (AI/Deep Learning): These were great at removing the noise, but they were too aggressive. They cleaned the signal so well that they accidentally erased the important brain data too. It's like using a sledgehammer to clean a watch; you get rid of the dirt, but you break the gears.

3. The Secret Weapon: Using Motion to Fix the Brain
Here is the coolest part. They realized that the Motion Sensors (IMU) could actually help fix the Brain Signals.

The Analogy: If you know exactly how hard the floor is shaking (from the motion sensor), you can mathematically subtract that shaking from the microphone recording.
The Result: By using the motion data to "cancel out" the noise in the brain data, they got much clearer signals than before. It's like Noise-Canceling Headphones, but for your brain.

🚀 Why This Matters (The Future)

This dataset is a huge step forward because it moves brain technology out of the lab and into the real world.

Better Health: Imagine a headset that helps a stroke patient relearn to walk while they are actually walking around their house, not just sitting in a clinic.
Smarter AI: By combining what the person is doing (video), moving (sensors), and thinking (brain), we can build AI that truly understands human behavior.
Real-World Use: It proves that we can eventually have "invisible" brain interfaces that work while you are running, talking, or driving, not just while you are frozen in a chair.

🏁 In a Nutshell

The WearBCI paper is like a field guide for the future of brain technology. It says: "We used to think you had to be perfectly still to read your brain. We were wrong. We collected a massive library of data showing how movement messes up the signal, and we found that by using motion sensors to help clean the noise, we can finally make brain-computer interfaces work in the real, messy, moving world."

1. Problem Statement

While Brain-Computer Interfaces (BCIs) hold promise for medical diagnostics, neurorehabilitation, and human-computer interaction, their real-world deployment is hindered by motion artifacts.

The Gap: Existing wearable BCI datasets are primarily collected in stationary, controlled laboratory settings (e.g., sitting or resting). They fail to account for the signal degradation caused by body movement, environmental variability, and muscle activity (EMG).
The Challenge: Wearable systems use dry electrodes with higher impedance, making them highly susceptible to motion-induced voltage jumps, electrode shifts, and overlapping EMG signals. Current denoising methods (both traditional and deep learning) lack robust benchmarks for these dynamic, real-world scenarios.

2. Methodology: The WearBCI Dataset

The authors introduce WearBCI, the first multimodal dataset designed to comprehensively evaluate wearable BCI signals under varying motion dynamics.

Data Collection

Participants: 36 healthy adults (ages 18–26).
Hardware:
- EEG: OpenBCI Cyton board with 8 dry electrodes (Fp1, Fp2, P7, P8, T3, T4, O1, O2) sampled at 250 Hz.
- IMU: 5 inertial measurement units (head, wrists, ankles) sampled at 100 Hz.
- Video: Egocentric video recorded at 30 FPS (during navigation only).
Synchronization: All streams synchronized via Network Time Protocol (NTP) with Unix timestamps.

Experimental Protocol (Three Motion Levels)

The study progresses from low to high motion complexity:

Static Baseline: Sitting still (eyes open/closed) to establish physiological references.
Body Movements: Five discrete actions (head shaking, nodding, eye movement, arm stretching, typing) to isolate specific artifact sources.
Walking: Continuous locomotion at three speeds (2, 3, 4 km/h).
Natural Navigation: A complex indoor scenario involving six sequential tasks (opening doors, greeting, obstacle avoidance, walking corridors, reading posters, knocking). This combines movement with environmental interaction.

Analysis Framework

Signal Characterization: Analysis of time-domain features, Power Spectral Density (PSD), and topographic maps to quantify how motion intensity and type affect different brain regions.
Benchmarking: Evaluation of six representative EEG enhancement methods:
- Traditional: Independent Component Analysis (ICA), Empirical Mode Decomposition (EMD), Artifact Subspace Reconstruction (ASR).
- Deep Learning: GAN-based (GCTNet), Transformer-based (EEGDNet), Diffusion-based (EEGDfus).
Metrics: Weighted PSD Distance ( $D_{PSD}$ ), Alpha Peak Consistency ( $D_{\alpha}$ ), EMG Index, and Brain IC Ratio (neural preservation).

3. Key Contributions

First Multimodal Motion Dataset: Introduction of a synchronized dataset (EEG, IMU, Video) from 36 subjects covering static, discrete movement, walking, and complex navigation.
Systematic Benchmarking: A comprehensive evaluation of how motion intensity degrades signal quality and how current denoising algorithms perform under these conditions.
Cross-Modal & Behavioral Insights: Two novel case studies demonstrating:
- Cross-Modal Enhancement: Using IMU data to assist EEG denoising.
- Multi-Dimensional Behavior Understanding: Using fused multimodal data to decode scene, action, and cognition simultaneously.

4. Key Results

Signal Degradation Analysis

Correlation: There is a strong linear correlation between IMU acceleration (motion intensity) and EEG signal variability (STD).
Spectral Impact: Motion causes broadband spectral elevation. While mild movement causes minor distortion, navigation and abrupt actions (e.g., obstacle avoidance) cause severe spectral shifts and time-domain fluctuations.
Regional Specificity: Different actions contaminate specific brain regions (e.g., arm stretching affects parietal/temporal areas; knocking affects frontal/occipital areas).

Benchmarking Denoising Methods

Traditional Methods (ICA, EMD, ASR): Performance degrades significantly as motion intensity increases. For example, ICA's spectral distortion ( $D_{PSD}$ ) jumped from ~16 (body movements) to ~50 (navigation). They fail to handle non-stationary signals.
Deep Learning Models: Achieved lower spectral distortion ( $D_{PSD} \approx 9.0$ ) across all conditions. However, they suffered from "over-cleaning," suppressing meaningful neural signals (low Brain IC Ratio < 0.35), indicating they remove neural components along with artifacts.
Conclusion: No single method currently solves the trade-off between artifact suppression and neural preservation in dynamic environments.

Case Study Findings

Cross-Modal Enhancement: A simple linear regression model using IMU data to predict and subtract artifacts from EEG improved performance in navigation scenes ( $D_{PSD}$ reduced to 10.8) and retained higher neural fidelity (Brain IC Ratio = 0.52) compared to traditional methods.
Behavior Understanding:
- Modality Strengths: Video was best for "Scene" (0.864), IMU for "Action" (0.649), and EEG for "Cognition" (0.451).
- Fusion: An optimized fusion strategy combining all modalities achieved the highest overall semantic alignment score (0.859), significantly outperforming single-modality approaches (Video-only: 0.622).

5. Significance and Future Impact

Real-World Deployment: WearBCI provides the necessary infrastructure to develop robust algorithms for everyday BCI use, moving beyond the "lab-only" paradigm.
Algorithmic Guidance: The results highlight that current deep learning models risk over-suppressing neural data, suggesting a need for new architectures that balance artifact removal with signal fidelity.
Multimodal Necessity: The study proves that relying solely on EEG is insufficient for real-world applications; integrating IMU and video is critical for both signal cleaning and holistic behavior understanding (e.g., monitoring attention in Alzheimer's patients or stroke rehabilitation).
Resource Availability: The dataset and code are publicly available, serving as a standard benchmark for the community to advance wearable BCI research.