Objective: To evaluate whether surface electromyography (sEMG) signals from the biceps can detect dyskinesia in patients with Parkinson’s disease (PD) using machine-learning methods.

Background: Levodopa-induced dyskinesias are a frequent complication of PD and significantly affect quality of life. In clinical practice, dyskinesia assessment relies mainly on observational rating scales, which may be subjective and limited to brief clinical visits. Current wearable approaches for dyskinesia monitoring mainly rely on accelerometer signals. Biosignals such as sEMG may provide complementary physiological information and support more objective monitoring of dyskinetic activity.

Method: Data were obtained from the MoMoPa-AM project. For this analysis, 20 patients with PD were selected: 10 with upper-limb dyskinesia (UDysRS part 3 right or left arm/shoulder > 1) and 10 age and sex matched patients without dyskinesia. Mean age was 65 ± 7.3 years and H&Y 2.6 ± 0.4. sEMG signals from both biceps were analyzed independently. Dyskinetic segments were identified using clinician UDysRS part 3 annotations and divided into 30-second windows. Equivalent windows were randomly extracted from matched non-dyskinetic patients to ensure class balance. A total of 1,796 windows were analyzed. Signal features were extracted to train a support vector machine classifier, which performs well for bi-classification problems. Model performance was evaluated using leave-one-group-out cross-validation. Performance was assessed both at window level and at patient level by aggregating window predictions.

Results: Patient-level classification showed good discrimination between dyskinetic and non-dyskinetic patients. Across all cross-validation folds, the model achieved a sensitivity of 70% and specificity of 90% (TP=7, FP=1, TN=9, FN=3), stable across group combinations. Feature selection improved classification performance compared with models using the full set of extracted features.

Conclusion: sEMG devices combined with machine learning shows potential for objective detection of dyskinesia in PD. While classification of short segments remains challenging due to the intermittent nature of dyskinetic activity, aggregating predictions across patients significantly enhances discrimination. These findings support the potential EMG-based monitoring as a complementary tool for dyskinesia assessment.

« Back to 2026 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/semg-for-dyskinesia-detection-in-parkinsons-disease-using-machine-learning/