Source-linked AI summary
Detecting Interbrain Synchronization in EEG Hyperscanning with MUSE-S EEG headban
Tarmo Lipping, Ahmad Sharif, Matin Beiramvand, Jari Turunen
TL;DR
The study asks whether consumer-level MUSE-S EEG can detect interbrain synchronization during hyperscanning. It analyzes paired participants playing Tetris, extracts spectral and cross-coherence features, and trains CNNs to distinguish gaming from relaxation. Full features achieved near-perfect discrimination, while inter-subject cross-coherence alone reached 94% training and 79% testing accuracy, supporting further investigation.
Problem
The study addresses whether a consumer-level MUSE-S EEG headband can detect interbrain synchronization in hyperscanning.
Method
Five pairs completed alternating relaxation and three two-person Tetris sessions, whose EEG features were classified with CNN models using full or inter-subject cross-coherence sets.
Results
Using only inter-stream cross-coherence, classification reached 94% training accuracy and about 75–80% testing accuracy, including 79% testing accuracy after 50 epochs.
Takeaways & Limitations
Inter-subject cross-coherence features still discriminated relaxation from gaming, indicating that the consumer EEG headband can detect some interaction between participants’ brain activity.
Abstract
from arXiv · showhide
In this study preliminary results on the classification of EEG hyperscanning data acquired using MUSE-S consumer-level EEG headband are presented. Five pairs of subjects were involved and the recording protocol contained three two-person tetris game sessions alternating with relaxation periods. The data were segmented and ten spectral and cross-coherence features were calculated. The features were arranged into feature matrices and Convolutional Neural Network model was trained to discriminate between the relaxation and gaming. Two different feature sets - the full set and a set containing only inter subject cross-coherence features were tested. The results indicate that using the full feature set, relaxation and gaming periods were perfectly discriminated. Using only inter-subject cross-coherence features 94 % and 79 % classification accuracy for training and testing data was obtained, respectively.
I. INTRODUCTION
The study investigates whether consumer-oriented EEG can detect interbrain synchronization during hyperscanning. It motivates EEG as a feasible real-life modality and presents a preliminary MUSE-S investigation.
- Hyperscanning simultaneously measures brain activity from multiple subjects and supports research on social interaction, teamwork, and collaboration.
- EEG is presented as the most feasible brain-imaging technique for real-life hyperscanning applications.
- Consumer-oriented EEG devices are increasingly used in real-life situations, including prior MUSE-S mental-workload classification studies.
- The present study preliminarily evaluates whether MUSE-S can detect interbrain synchronization and informs a more extensive experiment.
A. Recording setup
Five two-person hyperscanning recordings used MUSE-S headbands during alternating relaxation and two-person Tetris sessions. The protocol included cooperative, control-swapping, and competitive game conditions.
- Five two-person recordings involved five healthy adult volunteers using MUSE-S EEG headsets with textile electrodes.
- The analysis selected the AF7 and AF8 channels, while TP1 and TP2 were discarded because of excessive artifacts.
- Three Tetris game sessions alternated with 1-minute relaxation periods in each recording.
- The sessions represented cooperation, swapped control over separate wells, and competition between participants.
S EEG devices over Bluetooth. The MuseLSL2 software1
The recording setup used two laptops to capture synchronized MUSE-S streams and present the Tetris task to both participants. LabRecorder stored EEG and keystroke streams for later segmentation.
- The setup used Lab Streaming Layer streams and stored EEG together with recorded keystrokes in .xdf files.
- Laptop A ran LabRecorder, the Tetris game, keystroke recording, and MuseLSL2 for one participant’s data.
- Laptop A provided duplicate monitors and keyboards so the two participants could use the game setup.
- Laptop B ran MuseLSL2 to capture the other participant’s MUSE-S data.
B. Data Preprocessing and Feature Extraction
The data were synchronized, filtered, segmented into overlapping windows, and represented with spectral and coherence features. Full and inter-stream cross-coherence feature matrices were prepared for CNN classification.
- EEG signals sampled at 256 Hz were visualized, marked by session boundaries, and filtered from 0.5 to 42 Hz with a 6th-order Butterworth bandpass filter.
- The two data streams were resampled so their samples occurred at exactly the same timestamps.
- Sessions were divided into 10-second segments with 8-second overlap, producing 475 relaxation and 2161 gaming segments.
- Ten features comprised four scaled power spectral densities, two within-stream coherences, and four inter-stream coherences.
- Features were computed from 0 to 45 Hz in 0.5 Hz steps, forming 91 × 10 matrices that were augmented for CNN input.
- An inter-stream-only subset used four features and was augmented into 91 × 20 matrices for classification.
C. Classification
The study uses a CNN to classify augmented EEG feature matrices, with separate configurations for the full feature set and inter-stream cross-coherence features. Training used repeated runs with fixed data splits and balancing of the gaming segments.
- C. Classification: A CNN classifier was applied to augmented EEG feature matrices to identify patterns associated with the two conditions.The network included convolutional and fully connected layers with ReLU and sigmoid activations.
- C. Classification: The feature matrices represented either the full feature set or only inter-stream cross-coherence features.The two feature configurations corresponded to the alternatives evaluated by the classifier.
- C. Classification: The full-feature configuration used batch size 32 and learning rate 0.00009, whereas the inter-stream configuration used batch size 16 and learning rate 0.0002.Both configurations used Adam optimization and binary cross-entropy loss.
- C. Classification: 70% of the data were randomly assigned to training and 30% to testing across 50 runs of 50 epochs.Gaming samples were randomly downsampled to match the 475 relaxation feature matrices in each run.
III. RESULTS AND DISCUSSION
The CNN distinguished relaxation from gaming using the full feature set and also achieved substantial classification accuracy using only inter-stream cross-coherence features. The latter configuration learned more slowly and reached lower testing accuracy than training accuracy.
- III. RESULTS AND DISCUSSION: The full-feature model learned to discriminate relaxation and gaming during approximately the first few epochs, despite a learning rate of 0.00009.Training loss and classification accuracy were aggregated over 50 runs for the reported comparison.
- III. RESULTS AND DISCUSSION: The full feature set produced nearly perfect discrimination between relaxation and gaming despite data acquisition with a consumer-oriented wireless EEG device.The authors attribute part of this separability to the differing eye states between conditions.
- III. RESULTS AND DISCUSSION: 94% training accuracy was reached after 50 epochs using only inter-stream cross-coherence features.Testing accuracy became limited at about 75–80% after 20 epochs.
- III. RESULTS AND DISCUSSION: 79% accuracy using only interbrain features and consumer-oriented EEG devices was considered noteworthy.The authors also noted information leakage between subjects because there were only five subjects and random cross-validation was used.
IV. CONCLUSIONS
Using consumer-oriented MUSE-S EEG headbands, the study discriminated relaxation from gaming with spectral and cross-coherence features. Inter-subject cross-coherence alone also supported discrimination, indicating detectable interaction between participants’ brain activity, while broader recordings are needed.
- Five hyperscanning EEG recordings used CNN models to discriminate relaxation and gaming during two-player Tetris sessions.The analysis used power spectral density and cross-coherence features.
- Spectral and cross-coherence features enabled easy discrimination of relaxation and gaming with a low-complexity consumer-oriented EEG headband using prefrontal channels.
- Inter-subject cross-coherence features alone still discriminated the two brain states, indicating that the headband detected some interaction between subjects’ brain activity.
- More extensive hyperscanning recordings are needed, and future experiments will better control collaborative and competitive tasks.The study did not promisingly discriminate the different game types.