Feature-Level Fusion of Surface Electromyography and Mechanomyography Signals for MVC-Normalized Shoulder Abduction Force-Level Classification in Healthy Adults.

BACKGROUND

Accurate recognition of upper-limb force levels is important for wearable movement monitoring and rehabilitation engineering, yet the value of combining surface electromyography (sEMG) and mechanomyography (MMG) for shoulder force classification remains incompletely characterized.

METHODS

Ten healthy adults performed right shoulder abduction at four maximum voluntary contraction (MVC)-normalized force levels of approximately 10%, 30%, 60%, and 90% MVC. Signals were synchronously collected from the middle deltoid at 1000 Hz and segmented using 500 ms windows with a 150 ms stride. The evaluated classifiers were logistic regression (LR), k-nearest neighbors (KNN), decision tree (DT), support vector machine with a radial basis function kernel (SVM-RBF), random forest (RF), extremely randomized trees (ET), histogram-based gradient boosting decision tree (HGBDT), and multi-layer perceptron (MLP). Models were evaluated using group-aware five-fold cross-validation at the action-trial level.

RESULTS

The dataset contained 12,231 windows from 30 action-trial groups. HGBDT achieved the best performance, with an accuracy of 0.904±0.023, macro-F1 score of 0.911±0.018, quadratic weighted Cohen's kappa of 0.910±0.038, and mean absolute grade error of 0.137±0.042. Fusion increased macro-F1 from 0.821±0.030 for sEMG-only and 0.771±0.015 for MMG-only to 0.911±0.018.

CONCLUSIONS

These internally validated findings support the complementary value of sEMG and MMG for MVC-normalized shoulder force-level classification in healthy adults. Subject-independent and patient-level validation is required before clinical rehabilitation use.