TL;DR
This study demonstrates that a regression-based approach for controlling prosthetic hands, enabling simultaneous and proportional activation of two functions, can be effectively transferred from the lab to daily life. A single training session was sufficient for a transradial amputee to outperform their conventional prosthesis in functional tests, with performance further improving over an 8-week home trial.
What Problem Does This Paper Address?
Traditional myoelectric hand prostheses, while offering a degree of restoration, are often limited by simple control schemes. Many rely on just two EMG electrodes to control a single function (like opening or closing the hand) at a time, requiring cumbersome mode-switching techniques for additional functions. While newer classification-based controllers allow access to more functions, they typically still operate sequentially. This limits the naturalness and efficiency of prosthetic use, especially for complex daily tasks. The authors previously showed in lab tests that a regression-based mapping of EMG signals could enable simultaneous control of two prosthetic functions (e.g., hand grasp and wrist rotation) with independent velocity control, achieving high reliability. The core problem this paper addresses is whether this advanced, high-fidelity control can translate from controlled laboratory settings to the unpredictable and dynamic environment of daily life.
Methods
The study employed a linear regression (LR) model to map eight-dimensional EMG signal envelopes to a two-dimensional control signal for a prosthetic hand. This approach differs from classification by directly estimating the desired output for each degree of freedom (DOF) simultaneously, rather than assigning the EMG pattern to a predefined class.
The participant, a transradial amputee with extensive experience using conventional myoelectric prostheses, was fitted with a research prosthesis equipped with eight EMG electrodes and a customized controller. The LR model was trained using a brief session (less than 1 minute) where the participant performed four specific phantom-limb motions. The training data was used to establish a linear mapping (ŷ = WTx, where W = (XᵀX)⁻¹XᵀY) from EMG envelopes (X) to control signals (Y).
Crucially, no retraining of the LR model was required during the entire 8-week home trial. However, post-processing thresholds were adjusted in the first two weeks to fine-tune speed and reduce unintended activations. The research prosthesis was compared against the participant's own commercially available Michelangelo hand, controlled with a conventional slope-control and co-contraction (CC) approach.
Datasets
The study did not introduce new benchmark datasets. Instead, it utilized:
- Participant's own EMG data: Collected during the training session for the regression model.
- Functional test data: Collected during laboratory sessions using standardized tests.
- Questionnaire data: Collected from the participant throughout the home phase.
- Data log: Continuous recording of EMG envelopes and motion data from the research prosthesis during the home phase.
Key Insights and Ideas
- Lab-to-Life Transferability: The most significant insight is that advanced, simultaneous, and proportional prosthetic control, achieved through regression, is not confined to laboratory settings. It can be successfully implemented and utilized in the complex, uncontrolled environment of daily life.
- Minimal Training Requirement: The fact that a single, brief training session was sufficient for the LR model and no retraining was needed throughout the 8-week study highlights the robustness and efficiency of this control approach. This is a major advantage over classification-based methods that often require frequent retraining.
- Intuitive Control and User Satisfaction: The participant reported the LR control to be highly intuitive, allowing for seamless switching between control modes and a more natural experience compared to his conventional prosthesis. This led to high satisfaction and a preference for the LR system.
- Naturalness of Multi-DOF Movements: The observed pattern of multi-DOF motions (e.g., combining grasp and rotation) suggests that the LR control facilitates more natural, preparatory movements for grasping, which are difficult to achieve with sequential control.
- Potential for Improved Functionality: The study's findings strongly suggest that regression-based control can lead to significant improvements in functional performance, as evidenced by the participant's superior performance in standardized tests.
Key Results
- Functional Tests:
- Box-And-Blocks Test: The participant performed significantly better with the LR prosthesis than with his conventional prosthesis (CC) even before the home phase. This advantage was maintained after the home phase.
- Clothespin-Relocation Test: While performance was similar initially, after the 8-week home phase, the LR control showed significant improvement, with no dropped pins compared to some dropped pins with CC.
- SHAP Test: The participant achieved a higher SHAP score (a measure of overall hand function) with the LR prosthesis (53) compared to CC (45) at the beginning of the study, further improving to 62 after the home phase.
- Daily Wear Time: The participant wore the research prosthesis for an average of 5-7 hours per day in the initial weeks, decreasing to 1-4 hours per day in the latter half. This decrease was attributed to the more comfortable socket and preferred aesthetic of his conventional prosthesis, rather than dissatisfaction with the control itself.
- Movement Activity: After initial adjustments, the number of multi-DOF motions decreased significantly, while single-DOF motions (particularly open/close) increased. This suggests a refinement of control and potentially a reduction in unintended activations.
- Subjective Assessment: The participant consistently rated the reliability and naturalness of the LR control higher than CC. He also perceived the LR-controlled prosthesis as more "own hand" and reported fewer dropped items with LR.
Limitations and Future Work
- Single Case Study: The study is a case study of a single participant. While promising, the results may not generalize to all users with transradial amputations, as individual muscle signal characteristics and adaptation abilities vary.
- Socket Comfort: The research prosthesis socket, while professionally made, was noted as less comfortable than the participant's own due to the pressure required for the eight electrode modules. This may have influenced daily wear time. The authors suggest exploring conductive inserts for improved comfort in future designs.
- Comparison Bias: The comparison involved two different prostheses, with the Michelangelo hand having potentially faster maximum speeds. The authors acknowledge this potential bias.
- Unintended Activations: While the participant reported no false activations after threshold adjustments, the study acknowledges that it's not definitively proven that all recorded multi-DOF motions were intended.
- Future Work: The authors emphasize the need for further studies with a broader range of prosthetic users to confirm the clinical relevance and broad applicability of regression-based control. Investigating the impact of different socket designs and exploring methods to further improve comfort are also important future directions. Additionally, comparing the frequency of motions per hour for CC and LR would provide further insights.
Summary updated