Supervised and Semi - Supervised Machine Learning Networks applied for control of a Lower - Limb Exoskeleton

Abstract

Brain–machine interfaces (BMI) for lower-limb exoskeletons are a state-of-the-art neurorehabilitation modality. They decode electroencephalographic (EEG) recordings during motor imagery (MI)—the mental rehearsal of movement—to infer intent and drive exoskeleton control. Yet MI decoding suffers from low signal-to-noise ratio, EEG non-stationarity, and high inter-trial/subject variability. Conventional machine-learning classifiers further struggle with limited training data and overfitting, undermining real-time robustness. In this preliminary, offline study on a single subject, a novel semi-supervised MI-classification network is implemented that includes an L2-normalized autoencoder with dual reconstruction and classification branches—that, to our knowledge, is the first correctly tailored for closed-loop lower-limb exoskeleton control. This method is compared against four supervised approaches using a hybrid feature-extraction pipeline capturing spectral, spatial, and temporal EEG dynamics. Supervised models were evaluated via leave-one-out cross-validation, while the semi-supervised framework’s latent representations were examined with K-means clustering and t-Stochastic Neighbour Embeddings (t-SNE). Event-based false-positive (FPR) and true-positive ratios (TPR) served as comparative metrics. All approaches achieved 61–67 % accuracy, with the semi-supervised network showing a lower FPR—suggesting its promise for more robust, data-efficient BMI-driven exoskeleton control.