Machine Learning-Based EEG Classification for Controlling a Lower-Body Robotic Exoskeleton

Brain–computer interface (BCI) research for lower-limb assistance has advanced rapidly; however, many existing studies focus on upper-limb or binary tasks and rely on large multiuser datasets or deep neural models requiring substantial training data. This article presents a data-efficient EEG-based framework for multiclass lower-body movement recognition and its integration with a robotic exoskeleton. Electroencephalogram (EEG) data were collected using a 14-channel EMOTIV EPOC X headset from 15 participants (ages 23–78, 33% women) performing six periodic lower-body movements: resting stand, resting sit, walking, standing–sitting transition, spin left, and spin right. A single-subject dataset (180 recordings) was used for user-specific calibration, while the complete multisubject dataset (600 recordings) was used to evaluate robustness under intersubject variability. Signals were segmented into 2-s windows and processed using a discrete wavelet transform (DWT) with the bior4.4 wavelet at three decomposition levels. Statistical features (energy, entropy, mean, and standard deviation) were classified using a gradient boosted tree (GBT) model. In the single-subject calibration setting, the model achieved 0.742 ± 0.035 accuracy across six classes. In the multisubject evaluation, DWT-based classification achieved 0.718 ± 0.025, outperforming a random forest baseline (0.696 ± 0.022) and a lightweight 1-D convolutional neural network (CNN) trained on raw EEG signals (0.549 ± 0.036). Hybrid feature integration combining DWT, band power, and filter bank common spatial patterns (CSPs) further improved multisubject performance to 0.748 ± 0.018, increasing to 0.759 ± 0.020 with 90% overlapping windows. Reduced-class experiments achieved 0.895 ± 0.019 for three movements and 0.936 ± 0.015 for binary classification. The trained classifier was integrated with a six-degree-of-freedom servo-actuated exoskeleton driven by a Hopf oscillator-based central pattern generator (CPG), translating EEG-derived movement states into coordinated actuation commands. These results demonstrate that structured wavelet-based feature extraction combined with ensemble learning provides a computationally efficient and physiologically grounded strategy for adaptive lower-limb neuro-controlled assistive systems.