Lightweight Test-Time Adaptation for EMG-Based Gesture Recognition

TL;DR

This work tackles the problem of non-stationary EMG signals that degrade cross-day gesture decoding by introducing a lightweight Test-Time Adaptation framework built on a Temporal Convolutional Network. It validates three deployment-ready strategies—causal Adaptive Batch Normalization, replay-regularized statistical alignment via LoRA, and meta-learning with LoRA in a MAML framework—to bridge the inter-session accuracy gap with minimal overhead on resource-constrained wearables. Across NinaPro DB6, the methods show distinct strengths: AdaBN offers low-cost, real-time recalibration; distributional alignment with DER provides stability under limited data; and few-shot meta-learning delivers rapid gains when labels are available, with manageable compute and memory costs. The findings support a practical, hybrid deployment path for robust, long-term myoelectric control in wearable devices, potentially integrating privacy-preserving updates such as federated learning to personalize models without centralizing raw EMG data.

Abstract

Reliable long-term decoding of surface electromyography (EMG) is hindered by signal drift caused by electrode shifts, muscle fatigue, and posture changes. While state-of-the-art models achieve high intra-session accuracy, their performance often degrades sharply. Existing solutions typically demand large datasets or high-compute pipelines that are impractical for energy-efficient wearables. We propose a lightweight framework for Test-Time Adaptation (TTA) using a Temporal Convolutional Network (TCN) backbone. We introduce three deployment-ready strategies: (i) causal adaptive batch normalization for real-time statistical alignment; (ii) a Gaussian Mixture Model (GMM) alignment with experience replay to prevent forgetting; and (iii) meta-learning for rapid, few-shot calibration. Evaluated on the NinaPro DB6 multi-session dataset, our framework significantly bridges the inter-session accuracy gap with minimal overhead. Our results show that experience-replay updates yield superior stability under limited data, while meta-learning achieves competitive performance in one- and two-shot regimes using only a fraction of the data required by current benchmarks. This work establishes a path toward robust, "plug-and-play" myoelectric control for long-term prosthetic use.

Figures (5)

• Figure 1: Overview of the proposed framework for inter-session adaptation in surface EMG decoding. (a) 1D causal convolution and batch normalization applied over time, channels, and batch dimensions. (b) TCN architecture with $M$ residual causal-convolution blocks and a linear projection layer. (c) NinaPro DB6 setup: seven grasping gestures recorded using 14 double-differential EMG electrodes around the forearm. (d) Inter- vs. intra-session performance across training sessions $k\in\{2,3,5,7\}$, showing persistent cross-day domain shift. (e) Benchmark on NinaPro DB6 comparing accuracy and model size, highlighting the competitiveness of the lightweight TCN (0.05M parameters).
• Figure 2: Performance of the causal variant of batch normalization. (a,b) Batch-normalization adaptation: evolution of error rate on the current session (a) and on all other sessions (b) as a function of the mixing coefficient between old and new statistics. Results are reported for 2, 4, 7, and 14 gesture repetitions used for adaptation. (c) illustrates the shift of the UMAP projections across the adaptation strengths. (d,e) Online adaptation performance: evolution of the error rate over time for two representative subjects with different colors representing different adaptation speeds $\beta$.
• Figure 3: Performance of the statistical alignment strategy across sessions. (a,b) GMM statistical loss with DER alignment scheme's evolution of error rate on the current session (a) and on the other sessions (b) with respect to how big a buffer is required. (c) shows the performance of the batch-normalization adaptation and GMM adaptation without a DER component, while (d) compares the performance of the Covariance and GMM approaches with a DER component.
• Figure 4: Performance of Meta-learning across sessions. (a,b) illustrate the evolution of the error rate of the current session (a) and other sessions (b) across adaptation steps (x-axis) without DER and with respect to buffer size. The same data points are used for the different steps of the adaptation, as in few-shot learning settings. A comparative bar plot is made in (c) to display the effect of buffer size and DER on the final accuracy. Finally, (d) showcases the left-out session error rate when using the DER.
• Figure 5: Relative accuracy improvement of each test-time adaptation approach over the baseline with respect to the data buffer size, measured in seconds.