Comparative Analysis of Task-Specific and Combined Upper-Limb EMG Features for Early Parkinson’s Disease Classification

Abstract Early-stage Parkinson’s disease (PD) presents motor impairments that are difficult to detect clinically. Surface EMG (sEMG) offers an objective alternative, yet many studies rely on non-standardized tasks and provide limited task- and symptom-specific interpretability. This study analyzes sEMG recorded during standardized MDS-UPDRS-III upper-limb tasks—pronation–supination and postural tremor—performed by individuals with early-stage PD (Hoehn and Yahr 1–3, n=31) and healthy controls (n=30). Time-, frequency-, and nonlinear features were extracted and evaluated using a two-stage framework combining filter-based ranking and wrapper-based methods to support feature selection across multiple classifiers and interpretability. Pronation–supination showed the strongest single-task discrimination (balanced accuracy 0.79±0.181), driven by rhythm and nonlinear features reflecting impaired rhythmicity, reduced neuromuscular complexity, and unstable muscle deactivation, consistent with bradykinesia and rigidity. The postural tremor task highlighted tremor-specific spectral changes and reduced signal complexity during sustained posture (balanced accuracy 0.75±0.18), capturing low-frequency oscillations typical of PD tremor. Combining both tasks further improved classification without increasing feature dimensionality (balanced accuracy 0.83±0.186), indicating complementary diagnostic information. Filter-guided selection enhanced robustness and consistency across models. Beyond classification, this study highlights the value of interpretable, task-aligned motor quantification, showing that standardized clinical movements combined with targeted sEMG analysis can support explainable assessment of early-stage PD motor symptoms.