EMG-Based reaching prediction for upper limb rehabilitation: a systematic analysis of factors affecting the classification accuracy

Hybrid systems are increasingly used in clinical rehabilitation to restore upper limb functions, promoting motor relearning by linking voluntary intention with physical action. In this context, accurate classification of the intended movement direction, decoding residual muscle activity, can help optimize controller design. In this study, 22 healthy participants performed upper-limb reaching movements towards 9 spatial targets. Bipolar EMG signals were recorded from 10 muscles, and time- and frequency-domain features were extracted using time windows from 50ms to 1000ms. Three machine learning models were tested under two validation strategies: Leave-One-Subject-Out (LOSO) and holdout (80-20 split, 10-fold). We systematically investigated the impact of five factors on classification accuracy: number of muscles, feature number, time window length, classifier type, and validation method. Results showed that increasing the number of muscles significantly improved accuracy, while adding more than two features provided negligible gains. Long time windows enhance performance but may limit the use in real-time applications. With holdout validation, features extracted up to 350ms after the movement onset from five muscles reached ∼65% accuracy. Conversely, with LOSO validation, accuracy never exceeded ∼30%, revealing significant inter-subject variability. These findings highlight the need for subject-specific calibration to ensure robust, real-world applicability and provide practical indications for optimizing EMG-based rehabilitation systems.