This study, recently published by researchers at North-West University, systematically explores how feature extraction choices—specifically window size, step size, and feature type—impact the performance of Long Short-Term Memory (LSTM) networks for continuous finger-joint angle estimation. By balancing regression accuracy with the strict latency and memory constraints of embedded microcontrollers, the authors provide a roadmap for more responsive and intuitive myoelectric prostheses.
TL;DR
The paper systematically optimizes feature extraction for LSTM-based finger-joint angle regression, finding that a 300 ms window with a 100 ms step using Mean Absolute Value (MAV) or Root Mean Squared (RMS) features offers the best balance of accuracy and real-time feasibility. Their optimized approach achieves a test Pearson Correlation Coefficient (PCC) of 0.779, outperforming more complex architectures while remaining lightweight enough for deployment on standard prosthetic microcontrollers.
What Problem Does This Paper Address?
Most current myoelectric prostheses rely on gesture classification (e.g., "close hand" vs. "point finger"). While functional, this lacks the fluidity of natural movement. Researchers are shifting toward continuous joint angle regression, which allows a prosthesis to mirror the user's intended finger positions in real-time.
However, moving from classification to regression introduces two major challenges:
- Latency: To feel "natural," the delay between muscle contraction and prosthetic movement must stay below ~125 ms.
- Resource Constraints: Prosthetic controllers are typically embedded microcontrollers (MCUs) with limited memory (SRAM) and processing power.
While deep learning models like LSTMs are well-suited for this time-series data, the "pre-processing" stage—how we slice the raw muscle signals (sEMG) into windows and which mathematical features we extract—is often treated as a secondary concern. This paper argues that these design choices are critical and interact in non-trivial ways.
Methods
The authors employed a Long Short-Term Memory (LSTM) architecture, chosen for its ability to model temporal dependencies in sequences while remaining more lightweight than modern attention-based or heavy convolutional models.
The core of the experiment was a systematic sweep across three dimensions:
- Window Sizes: 300, 550, 800, 1000, and 1200 ms.
- Step Sizes: 10, 50, and 100 ms (constrained by the 125 ms latency limit).
- Features: Three Time-Domain (TD) features (RMS, MAV, Variance) and three Frequency-Domain (FD) features (Mean Frequency, Median Frequency, Mean Power).
A key methodological contribution is their rigorous Hyperparameter (HP) Tuning Protocol. They utilized two protocols: Protocol A (a light sweep for initial exploration) and Protocol B (an extensive sweep for final evaluation). This ensures that performance gains are attributed to the feature extraction parameters rather than lucky weight initializations.
Datasets
The study utilized the NinaPro DB2 dataset, a benchmark in the field.
- Participants: 40 subjects (29 male, 11 female).
- Hardware: 12-channel Delsys Trigno Wireless sEMG system (2 kHz sampling).
- Ground Truth: Finger-joint angles were captured using a 22-sensor CyberGlove II.
The researchers focused on Exercise C, which involves complex grasping motions, and predicted the first two joints of each finger.
Key Insights and Ideas
- Time-Domain Dominance: Simple time-domain features like MAV and RMS consistently outperformed frequency-domain features. This is a win for embedded systems, as TD features are computationally cheaper and require significantly less memory (192 bytes vs. up to 125 KB for FD features).
- The 300ms/100ms "Sweet Spot": While smaller step sizes usually improve performance in classification, the authors found that for 300 ms windows in regression, a larger 100 ms step actually performed best.
- Latency is Manageable: Even with the most complex feature/window combinations, the total time for feature extraction and inference was under 5 ms. This leaves ample "budget" (over 100 ms) for data acquisition and physical motor actuation within the 125 ms real-time limit.
- Joint Tuning is Mandatory: The interaction between window size and step size is non-linear. You cannot optimize one and then the other; they must be tuned as a pair to reach peak performance.
Key Results
The optimized LSTM models achieved a test PCC of 0.779 ± 0.080 using the MAV feature.
When compared to previous work (Long et al., 2023), this approach outperformed both the baseline LSTM and a more complex "LSTA-conv" model in terms of $R^2$ and PCC metrics. Crucially, the authors demonstrated that all tested configurations are feasible for deployment on an MCU with at least 128 KB of SRAM, which describes most modern ARM Cortex-M class processors used in robotics.
Limitations and Future Work
- Subject Specificity: The study noted that while the selected subset of 6-12 subjects was representative, individual differences in sEMG signals mean that feature choices might still need to be recalibrated for different users or different types of movements.
- Window Size Lower Bound: The study stopped at 300 ms because smaller windows became too noisy. Future work could investigate advanced filtering or denoising techniques to push window sizes even lower, potentially reducing initial latency further.
- Feature Fusion: The authors tested features individually. Combining multiple features (e.g., MAV + Mean Power) might yield higher accuracy at the cost of increased computational load, a trade-off worth exploring for higher-end prosthetic controllers.
Summary updated