Bridging the Performance Gap: The Impact of Grasping Behavior and Riemannian Geometry on Prosthetic Intent...

The control of myoelectric prostheses is persistently hindered by the poor translation of offline intent recognition accuracy to actual online performance. This study investigates the underlying mechanisms of this offline-online gap from a grasping behavior perspective and proposes a Riemannian geometry-based signal processing scheme.

We hypothesized that the feature distribution mismatch between models dominated by steady-state data and the requirements of online dynamic control is a primary factor in performance degradation. Using datasets from amputees and able-bodied subjects, we quantified the performance disparities of different training models across dynamic and steady-state phases, and evaluated the optimal training data composition for prosthetic control. Finally, based on this optimal data paradigm, we proposed a Riemannian geometry-based signal processing scheme to enhance feature separability.

Models neglecting grasping phase variations exhibited significant degradation in the dynamic phase, ranging from 5.85% to 18.16%. Conversely, models trained exclusively on dynamic data demonstrated superior intent recognition. Notably, the proposed method improved dynamic phase performance by up to 13.82%.

The mismatch between the steady-state distribution characteristics of training data and online dynamic task requirements is identified as a critical cause of the performance gap. The proposed scheme effectively enhances the model's decoding capability for dynamic signals by optimizing their geometric features in Riemannian space.

This work offers novel behavioral insights into the offline-online gap and provides an effective engineering solution for robust prosthetic systems.