Abstract
Using Machine Learning and Deep Learning to predict cognitive tasks from electroencephalography (EEG) signals has been a fast-developing area in Brain-Computer Interfaces (BCI). However, during the COVID-19 pandemic, data collection and analysis could be more challenging than before. This paper explored machine learning algorithms that can run efficiently on personal computers for BCI classification tasks. Also, we investigated a way to conduct such BCI experiments remotely via Zoom. The results showed that Random Forest and RBF SVM performed well for EEG classification tasks. The remote experiment during the pandemic yielded several challenges, and we discussed the possible solutions; nevertheless, we developed a protocol that grants non-experts who are interested a guideline for such data collection.
Introduction
Brain-Computer Interfaces (BCIs) have the potential to revolutionize medicine and neuroscience, yet the field is hampered by a significant scarcity of large-scale datasets. Unlike Computer Vision or Natural Language Processing, EEG data collection is traditionally expensive, requires clinical expertise, and is confined to controlled laboratory environments. The COVID-19 pandemic further restricted access to these facilities, highlighting the need for more flexible, remote-friendly data collection methods.
The BrainActivity1 framework addresses this "data drought" by investigating the feasibility of using consumer-grade hardware (under $300) and remote protocols to involve a broader audience, specifically college students, in EEG research.
Methods
The researchers developed a "closed-loop" workflow centered on an EEG Coach and an End-user.
Hardware and Remote Protocol
The study primarily utilized the Muse Headset, a popular non-invasive consumer device. To overcome pandemic-related restrictions, experiments were conducted via Zoom. A "coach" guided the user through hardware setup and experimental tasks remotely, providing real-time troubleshooting for signal quality issues.
Software and Analysis Pipeline
The framework includes a decision tree to assist users in selecting recording software (Mind Monitor vs. Muselab) based on their hardware and OS. Data was converted from proprietary .muse formats to CSV, cleaned using Matlab (preprocess.m), and filtered via Pandas.
For classification, the team tested several "classic" machine learning algorithms from the scikit-learn library, including Linear Classifiers, KNN, and Decision Trees. To optimize performance on standard personal computers, they focused on Random Forest and RBF Support Vector Machines (SVM), applying a Time Majority Voting (TMV) technique to refine final task predictions.
Datasets and Experimental Tasks
A pilot dataset was collected from college students using the "TCR Experiment" protocol. This involved five distinct cognitive tasks:
- T1: Think
- T2: Count
- T3: Recall
- T4: Breathe
- T5: Draw
The dataset includes six sessions per subject, designed to be processed by a student with a computer science background on a standard laptop in approximately two hours.
Results
The framework proved effective at classifying cognitive states using efficient models. Random Forest and RBF SVM were the most consistent performers. For Subject 1, the system achieved high precision across several tasks, notably 89% accuracy for "Think" (T1), 84% for "Draw" (T5), and 82% for "Breathe" (T4). Visualizations across six sessions validated that consumer-grade hardware can capture repeatable neural signatures for specific cognitive activities.
Discussion and Conclusion
While remote EEG collection is viable, the authors identified a "remote penalty." Virtual trials were significantly less efficient than in-person ones, primarily due to the difficulty of troubleshooting hardware issues—such as poor sensor contact behind the ear—without physical intervention.
Despite these hurdles, BrainActivity1 provides a scalable protocol for non-experts. Future work aims to build a "research/EEG coach-based student community" to foster peer-to-peer learning and troubleshooting, moving the field closer to the data scales seen in other AI sub-fields.
Summary updated