Source-linked AI summary
Electroencephalography Signal Analysis for Human Activities Classification: A Solution Based on Machine Learning and Motor Imagery
Tarciana C de Brito Guerra, Taline Nóbrega, Edgard Morya, Allan de M. Martins, Vicente A de Sousa
TL;DR
EEG-based BCI applications need reliable classification of motor activity, but EEG feature extraction and interpersonal variability remain challenging. This paper develops a Random Forest classifier for real and imagined movements using consumer-grade and research EEG systems. The method performs efficiently in same-subject, same-session conditions, while variability across times and individuals limits generalization.
Problem
Reliable extraction and classification of activity-related EEG features is difficult, and interpersonal variability complicates evaluation across subjects and recording conditions.
Method
The paper uses a supervised Random Forest system with hierarchical classification to identify movement region, laterality, and real versus imagined activity from EEG.
Results
Accuracies above 94% were achieved for Framework 1, while classification was not efficient in the other frameworks.
Takeaways & Limitations
Random Forest can classify real and imagined motor activities using consumer-grade EEG, especially when training and validation are performed for each use.
Takeaways & Limitations
Interpersonal and temporal EEG variability restricts classification across different collections, and Mindwave data limits performance in Frameworks 2 and 3.
Abstract
from arXiv · showhide
Electroencephalography (EEG) is a fundamental tool for understanding the brain's electrical activity related to human motor activities. Brain-Computer Interface (BCI) uses such electrical activity to develop assistive technologies, especially those directed at people with physical disabilities. However, extracting signal features and patterns is still complex, sometimes delegated to machine learning (ML) algorithms. Therefore, this work aims to develop a ML based on the Random Forest algorithm to classify EEG signals from subjects performing real and imagery motor activities. The interpretation and correct classification of EEG signals allow the development of tools controlled by cognitive processes. We evaluated our ML Random Forest algorithm using a consumer and a research-grade EEG system. Random Forest efficiently distinguishes imagery and real activities and defines the related body part, even with consumer-grade EEG. However, interpersonal variability of the EEG signals negatively affects the classification process.
1. Introduction
EEG-based BCI research targets assistive and rehabilitation technologies, but extracting reliable activity-related features remains difficult because EEG is noisy and complex. Machine learning is therefore used to interpret signals associated with real and imagined motor actions.
- BCIs can support accessibility, rehabilitation, and device control for people with motor disabilities and other users.
- EEG enables non-invasive, low-cost acquisition of brain activity with high temporal resolution for BCI applications.
- Reliable EEG interpretation requires extracting activity-related features and classifying signals across time, frequency, and spatial domains.
- EEG noise and complex pattern identification motivate machine-learning methods that can classify neural activity with less dependence on trained professionals.
- Motor imagery can support BCI protocols that do not depend on residual motor control, including neuroprosthesis control for people with paraplegia.
2. Related Works and Research Background
Prior work demonstrates machine-learning approaches for EEG and motor-imagery analysis, but the reviewed literature leaves limited evidence on Random Forest with consumer-grade EEG. The paper addresses this gap while examining task classification and subject-related signal variability.
- The reviewed studies used CNNs, signal decomposition, regression, neural networks, and other classifiers, while Random Forest was not established as the primary technique in this comparison.
- The literature review found no published comparison of consumer-grade EEG sensors with motor-imagery machine-learning algorithms.
- The paper asks whether Random Forest can classify real and imagined motor tasks, distinguish them at the FP1 electrode, and compare consumer-grade with research EEG devices.
- The study also investigates variability in EEG spectro-temporal and spatial characteristics across subjects.
3. Proposed Solution
The proposed system preprocesses EEG, derives statistical features, and applies a hierarchical Random Forest architecture to identify movement region, laterality, and whether movement was real or imagined. It evaluates this pipeline across consumer and research EEG acquisition settings.
- 3. Proposed Solution: The system acquires EEG during real or imagery tasks, preprocesses the signals, and compares classifier outputs with the performed trial.
- 3.1. Signal Preprocessing: A Butterworth band-pass filter attenuates powerline interference while retaining brain rhythms from delta through low gamma, and removes baseline drift before classification.
- 3. Proposed Solution: Random Forest classifies whether subjects performed real or imagined movements and identifies the involved limb among right and left hands and ankles.
- 3. Proposed Solution: The pipeline computes statistical moments of orders 1 through 10, splits data into training, validation, and test sets, and uses feature selection and grid search to configure Random Forests.
- 3.3. Implementation of the Proposed Solution: The hierarchical architecture uses Level A for hands versus ankles, Level B for right versus left within each region, and Level C for real versus imagery classification.
Frameworks’ Definition
The study defines three train-validation-test frameworks that progressively increase cross-subject variability to assess EEG classification difficulty. This design exposes a trade-off between classification ease and practical applicability.
- Frameworks’ Definition: The three frameworks progressively increase data variability across training, validation, and testing to evaluate classification difficulty between subjects.Framework 1 uses the same subject and collection; Framework 2 separates collections from one subject; Framework 3 tests on seven remaining subjects after selecting two reference subjects.
- Frameworks’ Definition: Framework 1 allocates 60%/20%/20% of data to training, validation, and testing, compared with 40%/10%/50% in Framework 2 and 18%/4%/78% in Framework 3.
- Frameworks’ Definition: The expected accuracy ranking is Framework 1 highest, followed by Framework 2 and Framework 3, because greater data variability tends to reduce accuracy.
- Frameworks’ Definition: Framework 1 requires retraining for each subject use, Framework 2 requires one training per subject, and Framework 3 supports one training for anyone at any time.The frameworks therefore trade easier classification against simpler application.
4. Experimental Setup and Measurement Protocol
The experiment compares simultaneous EEG recordings from a 16-channel research-grade V-AMP system and a single-channel consumer Mindwave device during real and imagined hand and ankle movements. Fourteen participants completed repeated 30-second tasks while a light sensor synchronized task timing with EEG acquisition.
- 4. Experimental Setup and Measurement Protocol: EEG was recorded simultaneously with a 16-channel V-AMP system and a single-channel consumer Mindwave device at 512 Hz.V-AMP used 16 active electrodes positioned according to the 10–20 system, whereas Mindwave used a dry forehead electrode and an earlobe reference.
- 4. Experimental Setup and Measurement Protocol: A screen presented each task instruction, while a white square appearing for 0.5 seconds at task onset and offset activated the light sensor.The synchronized light-sensor signal marked task boundaries in OpenVIBE and removed dependence on manual time stamping.
- 4. Experimental Setup and Measurement Protocol: Fourteen healthy participants performed real and imagined hand and ankle movements, repeating the 30-second task sequence twice with a rest period between recordings.Tasks included right- and left-hand opening and closing and right- and left-ankle flexion, in both real and imagined forms.
- 4. Experimental Setup and Measurement Protocol: Participants sat with their hands on their legs while following the task sequence displayed on a computer screen.
Data Selection
The study excluded recordings that violated task adherence, synchronization, or Mindwave streaming requirements. Consequently, nine of the fourteen participants remained for system analysis.
- Data Selection: Data were excluded when participants failed to follow instructions, synchronization triggers were missing, or Mindwave battery limitations impaired streaming.
- Data Selection: Nine of fourteen recorded participants satisfied the analysis conditions after five were excluded.The exclusions were intended to guarantee reliable results.
5. Results and Discussions
The results show that Random Forest classification can perform well with both research-grade V-AMP and low-cost Mindwave EEG, while variability across recordings and subjects remains a major constraint.
- Preliminary Analysis: EEG signals differed between people and across repeated collections by the same person, lacking a fixed visual pattern for a given activity.These variations help explain why machine-learning classification is needed and why temporal generalization remains difficult.
- Feature Selection: Around 100% accuracy was reached with approximately five selected inputs for hand-versus-ankle classification on both V-AMP and Mindwave.Feature selection reduced the number of mathematical operations while preserving satisfactory accuracy.
- Framework 1: Framework 1 achieved high and relatively stable accuracy across best- and worst-performing individuals because training, validation, and testing used the same subject’s data.The comparison also indicates that low-cost Mindwave can provide high classification accuracy despite its dry, single-channel design.
- Framework 2: Framework 2’s classification ability dropped considerably when training and testing used different collections from the same subject.The authors identify processing EEG data collected at different times as inefficient for the developed method.
- Framework 3: Framework 3 produced unsatisfactory and inconsistent sensor-dependent performance across subjects, with the best V-AMP result for imagined left-ankle movement and Mindwave result for imagined left-hand movement.The framework used average confusion matrices across all individuals, making the sensor difference explicit.
- Comparison Between Sensors: Across the frameworks, V-AMP offered superior data quality, whereas Mindwave’s channel-limited measurements still supported specific research applications.The devices generally reproduced consistent results despite their different hardware configurations.
6. Conclusions and Final Remarks
The Random Forest method classified real and imagery motor activities effectively, especially when training and testing used data from the same subject and collection. Performance declined across times and individuals, while consumer-grade EEG remained usable but less capable than research-grade equipment.
- The study used Random Forest to classify real and imagery activities and reported satisfactory convergence of its solution set.
- Random Forest efficiently distinguished and classified motor and imagery activities, particularly for data from the same subject and collection.The method was less effective when collections involved different times or different individuals.
- Above 94% accuracy was achieved in Framework 1, whereas classification was not efficient in the other frameworks.Framework 1 used the same data throughout training and classification; the results support exploring FP1 under this condition.
- Mindwave performed below V-AMP, but Framework 1 showed that consumer-grade EEG could still support EEG data-classification applications.Mindwave data restricted classification capacity in Frameworks 2 and 3.
- EEG data varied temporally among subjects, and interpersonal variability made classification across individuals especially complex.The authors identify this variability as a major difficulty and suggest future algorithmic improvements for cross-person classification.