EMG-based upper extremity motion intention detection with a two-layered multiple classifier system

Integrating electromyography (EMG)-based pattern recognition systems into the human–robot interface of upper extremity (UE) robotic exoskeletons can enhance patient engagement during rehabilitation. However, existing methods often fail to account for the multi-degree-of-freedom (mDOF) nature of the UE, transitional movement phases, and shared orientations among similar movements. To address these limitations, this paper proposes a two-layered multiple classifier system (MCS), composed of linear discriminant analysis (LDA) classifiers to classify five distinct UE movements that include both single-degree-of-freedom (sDOF) and activities-of-daily-living (ADL) motions. The first layer detects ambiguity in the input to assess prediction reliability; the second layer establishes motion intention by aggregating predictions from binary (MCS2) or ternary (MCS3) LDA classifiers trained on movement subsets. Using only the motion intention layer, the MCS significantly outperforms baseline multiclass LDA classifiers in prediction accuracy (MCS2: +2.6%, MCS3: +2.4%, p < 0 . 05 ). Adding the ambiguity detection layer yields added improvements (MCS2: +3.8%, MCS3: +8.7%, p < 0 . 05 ), with gains consistent across varying training data sizes ( p < 0 . 05 ). Overall, MCS3 with ambiguity detection demonstrates a superior ability to distinguish similar movements, identify transitional phases, and recognize UE motion intention more accurately, particularly during movement onset and progression. • Complex upper extremity motions have shared trajectories and orientations. • Conventional MCS does not consider orientation similarities during classification. • Ambiguity detection greatly reduces early misclassification. • MCS with ambiguity detection characterizes full complex upper extremity motion. • Ternary MCS outperforms binary MCS with combined ambiguity and intention detection.