TL;DR
This paper introduces "directional forgetting" for myoelectric control, a novel adaptation strategy that improves system stability and user performance by selectively discarding outdated training data. This approach leads to more robust and less sensitive prosthetic control compared to traditional methods.
What Problem Does This Paper Address?
Myoelectric prostheses, which use electrical signals from residual muscles to control artificial limbs, often require extensive initial calibration. This is because factors like muscle fatigue, electrode shifts, and perspiration can cause the EMG signals to change over time (non-stationarity), degrading control performance. While "co-adaptation" (where both the user and the system learn simultaneously) has shown promise, it's highly sensitive to how new data is weighted against old data. Classic "exponential forgetting" can lead to "estimator wind-up," where the system overfits to recent data, potentially causing users to lose control. This paper tackles the challenge of making co-adaptation more stable and reliable by developing a smarter way to forget old data.
Methods
The core of this work is the development of a Directional Forgetting (DF) scheme for Recursive Least Squares (RLS) algorithms, a common method for real-time adaptation. Unlike standard Exponential Forgetting (EF), which uniformly discounts all past data, DF selectively discards only the past information that is in the same "direction" as the new incoming data. This is achieved by decomposing the information matrix into components representing data to be kept and data to be discounted.
The researchers implemented this DF scheme within a co-adaptive myoelectric control framework. Participants performed virtual tasks, and if they failed to reach a target within a set time, the system would adapt its control model using either DF or EF. The performance was evaluated using metrics like Completion Rate (CR), Completion Time (CT), Path Efficiency (PE), and Throughput (TP).
Datasets
The study used data collected from five able-bodied participants performing real-time myoelectric tasks. No specific public datasets were used; instead, the researchers collected their own data through controlled experiments involving virtual cursor manipulation.
Key Insights and Ideas
- Selective Forgetting is Key: The most significant insight is that not all old data is equally "bad." Discarding data that is orthogonal to new data (i.e., representing different muscle activation patterns) is more beneficial than uniformly forgetting everything.
- Stability Over Aggression: Directional forgetting appears to strike a better balance between adapting to new user intentions and maintaining a stable, navigable control space, preventing the severe overfitting seen with aggressive exponential forgetting.
- Parameter Sensitivity Reduction: The study suggests that directional forgetting can alleviate the sensitivity to parameter selection (specifically, the forgetting factor), making the system more robust across a range of settings.
- User Learning Still Matters: While machine adaptation is crucial, the study also acknowledges the role of inherent user skill improvement, highlighting the need to differentiate between machine-driven adaptation and pure user learning.
Key Results
- Superior Performance: Directional forgetting with a forgetting factor of $\lambda = 0.995$ outperformed both exponential forgetting and unassisted user learning across multiple performance metrics (TP, CT, PE).
- Stable Control: Myoelectric control remained stable with directional forgetting across a range of forgetting factors, unlike exponential forgetting which could lead to instability at certain settings.
- Reduced Overfitting: Even with aggressive forgetting factors (e.g., $\lambda = 0.93$), directional forgetting did not induce complete loss of control, a problem reported with exponential forgetting at less aggressive settings.
- Weight Similarity: Analysis of model weights showed that RLS-EF induced more significant changes to the control model compared to RLS-DF, suggesting that DF leads to subtler, more stable updates.
Limitations and Future Work
- Limited Subject Pool: The study involved only five able-bodied participants, limiting the generalizability of the findings.
- Virtual Environment: The experiments were conducted in a virtual reality environment, which may not fully replicate the complexities and physical constraints of real-world prosthetic use.
- Arbitrary Time Limits: The 5-second time limit for triggering machine adaptation was arbitrary and might not perfectly align with user intent or exploration strategies.
- Future Directions: The authors suggest investigating the robustness of RLS-DF in non-stationary environments, exploring user-triggered adaptation, and, most importantly, conducting studies with actual limb-impaired participants using fully fitted prosthetic devices to validate the approach for clinical translation. They also note that the current approach assigns all input directly to the target, potentially ignoring true user intent, and future work could address this.
Summary updated