Abstract
Using Machine Learning and Deep Learning to predict cognitive tasks from electroencephalography (EEG) signals is a rapidly advancing field in Brain-Computer Interfaces (BCI). In contrast to the fields of computer vision and natural language processing, the data amount of these trials is still rather tiny. Developing a PC-based machine learning technique to increase the participation of non-expert end-users could help solve this data collection issue. We created a novel algorithm for machine learning called Time Majority Voting (TMV). In our experiment, TMV performed better than cutting-edge algorithms. It can operate efficiently on personal computers for classification tasks involving the BCI. These interpretable data also assisted end-users and researchers in comprehending EEG tests better.
Introduction
Brain-Computer Interface (BCI) technology faces a significant "small data" challenge. Unlike other AI domains, EEG studies often rely on limited cohorts. Furthermore, high-performing models typically demand clinical-grade hardware and substantial computing power, creating barriers for average users. This research addresses three primary hurdles for at-home EEG applications: signal noise from consumer-grade sensors, "mental outliers" caused by user distraction, and the high computational cost of deep learning models. The goal is to develop a classifier that is accurate, efficient on standard PCs, and interpretable enough to identify low-quality data.
Methods: The Time Majority Voting (TMV) Algorithm
The proposed Time Majority Voting (TMV) approach operates in two distinct phases to leverage temporal dependencies in brain activity—the principle that cognitive states persist over time rather than fluctuating instantaneously.
Phase 1: Selection Researchers benchmarked several standard algorithms, including Random Forest (RF), Support Vector Machines (SVM), and k-Nearest Neighbors (kNN). Random Forest and RBF SVM were identified as the top two performers and selected for the ensemble.
Phase 2: Ensemble Logic TMV uses a voting system where the "majority task" for a session is first determined by the best individual classifier (RF). If RF and SVM agree on a prediction, that result is accepted. In cases of disagreement, the algorithm defaults to the established "majority task" for that specific time block. This "smoothing" effect effectively filters out transient noise and mental outliers.
Data Preprocessing To ensure data quality, the first 30% of each task (the transition phase) was removed. Additionally, segments with sensor contact loss—identified by signal plateaus exceeding 1.4 seconds—were discarded.
Experimental Setup
The study utilized the Think-Count-Recall (TCR) dataset, involving 17 subjects (12 remained after cleaning). Participants performed five cognitive tasks: Think (T), Count (C), Recall (R), Breathe (B), and Draw (D). Each subject completed six 5-minute sessions over several weeks.
Results and Discussion
The TMV algorithm demonstrated a significant performance boost over standalone classifiers. While the best Phase 1 performer (Random Forest) achieved 55% accuracy, the full TMV pipeline—including noise filtering—reached an average accuracy of 80%.
Beyond accuracy, the system proved highly efficient. The total runtime for training and voting was approximately 74.3 seconds on a 2018 Macbook Pro, confirming its suitability for consumer hardware. The algorithm's interpretability also allowed researchers to diagnose specific sessions where sensor issues or user focus lapses occurred, such as Subject 3’s middle sessions.
Conclusion and Future Work
TMV demonstrates that clever ensembling of classical machine learning can rival complex deep learning models for BCI tasks while maintaining low computational overhead. However, the current approach relies on aggressive filtering, excluding sessions with low initial accuracy. Future work will focus on adapting TMV for real-time interactive environments and improving its robustness to noise without discarding significant portions of data.
Summary updated