Long-Term Upper-Limb Prosthesis Myocontrol via High-Density sEMG and Incremental Learning

TL;DR

The paper addresses the problem of long-term prosthetic myocontrol by tackling distribution shift in HD-sEMG signals and the limitations of low-density interfaces. It introduces a 64-channel dry HD-sEMG system and incremental learning via (RF-)RLSC to adapt to day-to-day variations, along with the DELTA dataset—a longitudinal HD-sEMG benchmark across 7 subjects and months. The authors demonstrate that incremental RF-RLSC maintains high accuracy under long-term shift, outperforming batch baselines, and provide fast predictions suitable for embedded prosthesis control. The work also releases the DELTA dataset and experimental code to foster research in robust, long-term myocontrol with wearable HD-sEMG sensors.

Abstract

Noninvasive human-machine interfaces such as surface electromyography (sEMG) have long been employed for controlling robotic prostheses. However, classical controllers are limited to few degrees of freedom (DoF). More recently, machine learning methods have been proposed to learn personalized controllers from user data. While promising, they often suffer from distribution shift during long-term usage, requiring costly model re-training. Moreover, most prosthetic sEMG sensors have low spatial density, which limits accuracy and the number of controllable motions. In this work, we address both challenges by introducing a novel myoelectric prosthetic system integrating a high density-sEMG (HD-sEMG) setup and incremental learning methods to accurately control 7 motions of the Hannes prosthesis. First, we present a newly designed, compact HD-sEMG interface equipped with 64 dry electrodes positioned over the forearm. Then, we introduce an efficient incremental learning system enabling model adaptation on a stream of data. We thoroughly analyze multiple learning algorithms across 7 subjects, including one with limb absence, and 6 sessions held in different days covering an extended period of several months. The size and time span of the collected data represent a relevant contribution for studying long-term myocontrol performance. Therefore, we release the DELTA dataset together with our experimental code.

Figures (8)

• Figure 1: Proposed system scheme. Pre-processed HD-sEMG signals $x_k$ at step $k$ are fed to the classifier $C_{\theta}$, which predicts the prosthesis gesture $\hat{y}_k$. $\hat{y}_k$ is filtered via time-window majority voting to improve control robustness. The dashed box shows the incremental model update process.
• Figure 2: Incremental RLSC Training.
• Figure 2: Optimal hyperparameters ($HP^{*}$) mean $\pm$ standard deviation for each participant. Statistics computed over the model selection results for each day of HD-sEMG acquisition. S: healthy subject, LDS: limb difference subject.
• Figure 3: DELTA dataset splits. Representation of 6 days of HD-sEMG acquisitions from a subject. 10 repetitions of each class are acquired each day. Batch setting: 2 repetitions for each class are included in the training set ($\mathcal{D}_{d=1,s}^{tr}$) for the first day, while the remaining constitute the test set. Incremental setting: same split for the first day, while for the following days 2 repetitions for each class are employed for incremental updates ($\mathcal{D}_{d,s}^{up}$) and the remaining 8 form the test set.
• Figure 4: 2D KPCA projection of the DELTA dataset for each of the 6 acquisition days (top), as well as for the entire dataset across days (bottom), for (a) a representative healthy subject and (b) the limb difference subject. Colors represent classes.