Unsupervised sEMG-Based Gait Anomalies Detection Using Autoencoder

Abstract

Gait Anomaly Detection is crucial for early diagnosis of neurodegenerative conditions and for monitoring rehabilitation progress. Surface electromyography provides a direct view into muscle synergy and motor intention, but the field suffers from a chronic scarcity of labeled pathological data. Traditional supervised methods, although accurate, rely on large, balanced datasets and often fail to generalize to unseen anomalies, limiting applicability in real-world clinical scenarios. An unsupervised framework based on a deep Autoencoder architecture is proposed. Instead of classifying anomalies, the model learns a robust representation of normal gait patterns through One-Class Classification, in a design that is inherently personalizable to each subject. Reconstruction errors are subsequently analyzed to identify deviations from normality, providing a sensitive measure of abnormal muscle coordination and motor patterns. The framework is validated on two distinct datasets: a controlled dataset of simulated anomalies, such as toe-walking, and a clinical dataset of real pathological gait patterns. Results demonstrate high detection accuracy: about 88% on real clinical knee pathologies and about 99% on the controlled setting of simulated toe-walking anomalies, the latter representing an easier, fully controlled scenario. These findings indicate that the framework effectively captures the muscle-activation synergies of normal gait and can detect deviations without requiring labeled pathological data. The unsupervised, reconstruction-based approach provides a practical, label-free solution for gait monitoring, with a normality model that can be adapted to the individual at deployment.