TL;DR
This paper introduces A-mode ultrasound as a novel, non-invasive sensing method to predict the walking kinematics of transfemoral amputees. By using an artificial neural network to map ultrasound signals from the residual limb to joint movements, the researchers achieved accurate predictions, paving the way for more intuitive and volitional control of powered prostheses.
What Problem Does This Paper Address?
Lower-limb powered prostheses hold immense potential for restoring mobility and independence to amputees. However, a critical challenge lies in reliably interpreting the user's intention to move. Current methods, like surface electromyography (EMG), often struggle with low signal quality and interference, limiting the prosthesis's responsiveness and the user's sense of control. While ultrasound offers better resolution and specificity, it hasn't been widely adopted for lower-limb prosthetics. This research aims to bridge this gap by exploring A-mode ultrasound as a viable sensing modality for predicting prosthesis kinematics, thereby enabling more direct volitional control.
Methods
The core of this study involves using A-mode ultrasound to capture signals from the residual limb muscles of transfemoral amputees during walking. Four ultrasound sensors were strategically placed on the anterior and posterior sides of the residual limb. These sensors captured 1D ultrasound signals, which were then processed to extract features. Specifically, the raw signals were offset, rectified, and a moving average convolution was applied to obtain the signal envelope. This envelope was then segmented into windows, and the average value within each window served as a feature.
These extracted ultrasound features were then fed into a regression neural network. The network was trained to map these features to the actual joint kinematics (knee and ankle position and velocity) of the prosthesis, which were simultaneously recorded using an IMU-based motion capture system. The neural network architecture consisted of multiple fully connected layers with ReLU activation functions, designed for efficient online prediction. The researchers trained individual models for each joint variable and tested them on data from different walking speeds to assess generalization.
Datasets
The study utilized data collected from nine transfemoral amputee subjects. For each subject, walking data was collected at two distinct speeds: a comfortable, self-selected speed (normal speed) and a faster self-selected speed. The data from the first two walking trials at normal speed were used to train the regression neural networks. The remaining two trials (one at normal speed and one at a faster speed) served as independent testing sets. No inter-subject models were trained due to variations in residual limb anatomy and sensor placement.
Key Insights and Ideas
- Ultrasound as a direct proxy for muscle deformation: The study demonstrates that A-mode ultrasound can effectively capture the subtle morphological deformations of residual limb muscles during walking. These deformations are directly linked to the user's intended movements.
- A-mode's potential for portability and affordability: Compared to B-mode ultrasound, A-mode is presented as a more compact and cost-effective solution, making it more practical for integration into prosthetic sockets.
- Neural networks excel at mapping complex signals: The use of a regression neural network proved effective in learning the intricate relationship between the 1D ultrasound signals and the multi-dimensional joint kinematics.
- Individualized models are key: The necessity of training subject-specific models highlights the variability in residual limb anatomy and sensor placement, emphasizing the need for personalized prosthetic control systems.
Key Results
The study achieved promising results in predicting transfemoral amputee prosthesis walking kinematics:
- Low Normalized RMSE: For the normal speed testing set, the mean normalized RMSE for knee position, knee velocity, ankle position, and ankle velocity were 9.0 ± 3.1%, 7.3 ± 1.6%, 8.3 ± 2.3%, and 10.0 ± 2.5%, respectively.
- Comparable performance across speeds: While the fast speed testing set showed slightly higher error rates (15.7 ± 5.4% for knee position, 12.6 ± 3.1% for knee velocity, 14.0 ± 3.9% for ankle position, and 13.1 ± 2.0% for ankle velocity), the predictions still closely matched the general shape of the recorded kinematics.
- Minimal delay with filtered predictions: A 2nd order Butterworth filter was applied to the predictions, resulting in smoother outputs with a minimal delay of 35.1 ± 5.8 ms, which is considered acceptable for online control.
- Accurate kinematic profiles: The filtered predictions visually matched the recorded kinematics, although a slight reduction in the range of motion was observed, particularly at faster walking speeds.
Limitations and Future Work
The authors acknowledge several limitations and suggest avenues for future research:
- Reduced range in predictions: The study observed that predicted kinematics consistently undershot actual values, leading to a smaller range of motion. Future work could explore additional A-mode ultrasound features or incorporate a gain adjustment to the model to address this.
- Generalization to varied speeds: While the model showed some generalization to faster walking speeds, the reduced range suggests that training with a wider variety of walking speeds might be necessary for improved model generalization.
- Sensor placement optimization: The current sensor placement was influenced by user comfort. Developing custom sockets that better accommodate the sensors could improve placement and reduce noise.
- Real-world application: The study was conducted in a controlled environment. Future research should focus on applying this methodology in real-world settings and investigating the impact of factors like fatigue.
- Integration into powered prostheses: The ultimate goal is to integrate this A-mode ultrasound sensing into actual powered prostheses and develop volitional controllers that leverage these predictions for direct user control.
Summary updated