Grouping sEMG Signals From Individuals With Neurological Injuries Using Clustering Methods

The contamination of surface electromyography (sEMG) data by artifacts is a critical issue, particularly when acquired from individuals with neurological injuries. This necessitates subjective and time-consuming visual inspection to verify signal utility. To address this, this work proposes an automated quality assessment method based on unsupervised learning to distinguish usable signals from those compromised by artifacts, such as electrode detachment or saturation. Features were extracted using the Time Series Feature Extraction Library (TSFEL), and the top 80 discriminative features across temporal, statistical, and spectral domains were selected via ANOVA. K-means and Agglomerative Clustering were employed to group the data, followed by an Isolation Forest stage to refine the selection of usable signals. Validation against an objective ground truth revealed the complexity of the task (ARI ≈ 0.20), yet the pipeline effectively retained 52% of the dataset as usable signals. Furthermore, a comparison between the retained injured signals and healthy controls demonstrated statistically significant differences (pd≈ 0.72), confirming that the automated approach preserves signals with distinct and biologically relevant spectral characteristics.