Small Data, Honest Metrics: Muse-2 EEG-Only Movement Detection

In rehabilitation scenarios where there are limited resources, and where there are no IMUs present, the ability to detect limb movement through only EEGs would allow inexpensive monitoring options. While there are consumer-grade EEGs that would be useful for detecting limb movement for health monitoring and assistive devices, the short recordings and limited channels make honest evaluations difficult, and are compounded by overlapping windowing which leads to leakage of information across train/test splits. Motivated by the question of how far leakage-safe evaluation would allow for the use of classical ML and time-domain features, we study an EEG-only right-arm detector on a public Muse-2 recording (TP9, AF7, AF8, TP10; 301 samples). There is a primary$30 / 15$(19 windows) followed by a dense$30/29$(272 windows), where we use set of features (statistics, Hjorth, TKEO, line length, derivative-RMS, laterality, and inter-channel coupling) to compare ridge-logistic, cost-sensitive SVM-RBF, and bagged trees. For leakage protection, we use blocked, repeated cross-validation with perfold standardization/PCA, light smoothing ($K=2$), and choosing operating points at Precision@Recall$\geqslant 0.80$. On the primary regime, ridge-logistic seems to have the best balance (BAcc 0.631; Recall 0.850; Specificity 0.412; F1 0.742), as SVM/bagging appear to trade specificity for higher recall. In the dense area, there is a significant advancement across all models in the evaluation of contiguous blocks, with SVM-RBF and bagged trees leading the way (BAcc$0.887 \pm 0.028 / 0.891 \pm$0.026; Precision$\approx 0.99$; Specificity$\approx 0.99$), while logistic models are still somewhat competitive. The results emphasize the value of a straightforward, reproducible pipeline using solely EEG and elucidate the reason for the dense windows in a leakage-safe evaluation.