Simplifying Prediction of Intended Grasp Type: Accelerometry Performs Comparably to Combined EMG-Accelerometry in Individuals With and Without Amputation

The adoption of active upper-limb prostheses with multiple degrees of freedom is largely lagging due to bulky designs and counterintuitive operation. Accurate gesture prediction with minimal sensors is key to enabling low-profile, user-friendly prosthetic devices. Wearable sensors, such as electromyography (EMG) and accelerometry (ACC) sensors, provide valuable signals for identifying patterns relating muscle activity and arm movement to specific gestures. This study investigates which sensor type (EMG or ACC) has the most valuable information to predict hand grasps and identifies the signal features contributing the most to grasp prediction performance. Using an open-source dataset, we trained two types of subject-specific classifiers (LDA & KNN) to predict 10 grasp types in 13 individuals with and 28 individuals without amputation. Having 4-fold cross-validation, LDA average accuracies using ACC only features (84.7%) were similar to combined ACC & EMG (88.3%) and much greater than with only EMG features (58.1%). Feature importance analysis showed that participants with amputation reached more than 80% accuracy using only three features, two of which were ACC-derived, while able-bodied participants required nine features, with greater reliance on EMG. These findings suggest that ACC is sufficient for robust grasp classification in individuals with amputation and can support simpler, more accessible prosthetic designs. Future work should focus on incorporating object and grip force detection alongside grasp recognition and testing model performance in real-time prosthetic control settings.