Design and Benchmark Evaluation of a Low-Cost Single-Channel EEG Framework for Attention–Relaxation Classification

Low-cost electroencephalography (EEG) systems offer a promising pathway toward accessible brain–computer interface technologies; however, their limited signal quality and hardware constraints present significant challenges for reliable cognitive-state classification. This paper presents the design and comparative evaluation of a low-cost single-channel EEG acquisition framework for Attention (AT) and Relaxation (RL) state classification using minimally processed EEG data. A BioAmp EXG sensor interfaced with an Arduino microcontroller was used to acquire a single 10-bit EEG signal stream from ten healthy participants. To intentionally investigate the performance limits of severely constrained EEG hardware, no frequency-domain feature extraction, temporal windowing, channel selection, spectral analysis, or advanced signal preprocessing techniques were employed. Instead, machine learning models were trained directly on raw EEG amplitude values. Five supervised machine learning algorithms, namely Fine Decision Tree, Quadratic Discriminant Analysis, Kernel Logistic Regression, Fine K-Nearest Neighbour, and Medium Neural Network, were trained and evaluated using MATLAB Classification Learner Toolbox. Performance was assessed using independent training and testing recordings together with accuracy, F1-scores, confusion matrices, ROC curves, Precision–Recall curves, training time, prediction speed, and computational cost. Experimental results indicate that Fine Decision Tree, Kernel Logistic Regression, and Medium Neural Network achieved the highest testing accuracy of approximately 56%, with macro and micro F1-scores in the range of 55–57%, while Quadratic Discriminant Analysis and Fine K-Nearest Neighbour exhibited near-chance performance. The modest classification accuracy reflects the absence of temporal, spectral, and spatial EEG features as well as the limitations associated with low-cost single-channel EEG acquisition. Nevertheless, the study provides a transparent benchmark of classifier behaviour under severe sensing and computational constraints and highlights the trade-offs between classification performance, computational efficiency, and hardware simplicity. The presented framework should therefore be interpreted as an exploratory baseline for future low-cost EEG research rather than as a validated real-time brain–computer interface system.