Designing a Model to Detect Beta Burst in EEG Using Nonlinear Dynamic Features Based on Machine Learning

Abstract

BackgroundBeta bursts are brief, transient increases in beta-band (13-30 Hz) EEG activity that play a key role in motor control, particularly in processes like movement initiation and inhibition. While most existing methods detect these bursts using simple amplitude thresholds, they often ignore variability in burst duration and task context. More advanced techniques exist but are computationally demanding, often opaque, and dependent on large datasets and strong modeling assumptions. ObjectivesThis study aims to develop an automated, machine learning-based approach to classify beta bursts using nonlinear dynamic features, considering both burst duration and task condition. MethodEEG data were collected from 26 healthy, right-handed participants during three motor tasks: (1) a right-hand isometric pinch grip at 10% maximum voluntary contraction (MVC), (2) rhythmic right-hand finger opening-closing in response to auditory cues (3-4 seconds), and (3) the same right-hand opening-closing task and concurrent with a steady left-hand isometric pinch grip at 10% MVC. Beta bursts were extracted from the left motor cortex, categorized by duration (short, medium, long), and time-locked to task events. From each burst, four nonlinear features, Fractal Dimension (FD), Wavelet Entropy (WE), Sample Entropy (SE), and Nonlinear Energy Operator (NEO) were calculated to train machine learning (ML) models. ResultsStatistical tests and feature selection revealed that FD, SE, and WE varied significantly with burst duration and task type, while NEO was more limited in sensitivity. ML models trained on these features achieved up to 91.1% validation and 85.7% test accuracy, especially when distinguishing bursts of different durations within the same task. ConclusionsThese findings suggest that beta bursts reflect structured, task-specific neural dynamics rather than random fluctuations. By using nonlinear features and ML, we introduce a scalable, interpretable framework for burst classification that outperforms traditional methods. This approach advances the understanding of transient beta activity and holds promise for real-time neural decoding in neuroscience and neurotechnology.