Abstract
Intelligent control strategies for active biomimetic prostheses could exploit the inter-joint coordination of limbs in human gait in order to mimic the functioning of a biological joint. A machine learning regression model could be employed to learn an input-output relationship between the coordinated limb motion in human gait and predict the motion of a particular limb/joint given the motion of other limbs/joints. Such a model could be potentially used as a controller for an intelligent prosthesis which aims to restore the functioning similar to an intact biological joint. For this, the model needs to be tailored for each user by learning the gait pattern specific to the user. The challenge of training such machine learning regression models in prosthetic control is that, the desired reference output cannot be obtained from an amputee due to the missing limb. In this study, we investigate the feasibility of using two different methods for training a random forest algorithm using incomplete amputee-specific data to predict the ankle kinematics and dynamics from hip, knee, and shank kinematics. First is an inter-subject approach which learns a generalized input-output relationship from a group of able-bodied individuals and then applies this generalized relationship to amputees. Second is a subject-specific approach which maps the amputee's inputs to a desired normative reference output calculated from able-bodied individuals. The subject-specific model outperformed the inter-subject model in predicting the ankle angle and moment in most cases and can be potentially used for devising a control strategy for an intelligent biomimetic ankle.
Introduction
Active biomimetic prostheses utilize motors to provide the torque and range of motion necessary for a natural gait. To control these devices effectively, the prosthesis must determine the appropriate ankle behavior based on the movement of the residual leg segments (hip, knee, and shank). While machine learning regression models are ideal for learning these complex, continuous input-output relationships, they face a significant "cold start" problem: how can a model be trained to predict an amputee's ankle movement when the biological ankle is missing?
Traditionally, researchers have relied on "inter-subject" models trained on able-bodied data. However, amputees often develop unique gait compensations that differ significantly from able-bodied norms, leading to suboptimal controller performance. This study explores whether a "subject-specific" approach—training on an individual's unique residual limb kinematics—can bridge this gap, even when the target ankle data is incomplete.
Methodology
The researchers employed Random Forest (RF) regression to predict ankle angle (kinematics) and ankle moment (dynamics). RF was selected for its ability to handle small datasets, lack of requirement for feature normalization, and computational efficiency suitable for real-time applications.
The study compared two primary training strategies:
- Inter-Subject Model: A generalized model trained on 24 able-bodied subjects.
- Modified Subject-Specific Model: A personalized model for each amputee. Because "true" ankle output was unavailable, the researchers used the median ankle trajectory of able-bodied subjects walking at matching speeds as the "target" for the amputee's specific hip, knee, and shank inputs.
The input features included sagittal plane angles and angular velocities of the hip, knee, and shank. The target outputs were ankle flexion-extension angle ($\theta_{ankle}$) and ankle moment ($\tau_{ankle}$). Data were sourced from a public dataset of 30 able-bodied volunteers and an experimental dataset from two unilateral transtibial amputees.
Key Insights and Results
The most significant finding was that the subject-specific model outperformed the inter-subject model, despite using normative target data. This suggests that capturing the unique "signature" of an individual's residual limb movement is more critical for prediction accuracy than having a perfect biological ground truth for the missing joint.
Performance Metrics
- Accuracy: The subject-specific model achieved high correlation with normative targets, with $R^2 > 0.97$ for ankle angle and $R^2 > 0.98$ for moment during normal walking.
- Error Reduction: This approach reduced Root Mean Square Error (RMSE) by up to three times compared to the inter-subject approach.
- Adaptability: The inter-subject model performed poorly on "Amputee 1," who showed significant gait deviations. However, the subject-specific model successfully adapted to these individual variations.
Feature Importance
Analysis revealed that shank orientation was the most critical predictor for ankle motion. Conversely, adding anthropometric data (height, weight, age) to inter-subject models actually decreased accuracy, likely by introducing inconsistent relationships that hindered generalization.
Discussion and Future Work
While promising, this study is limited by a small sample size of two amputees. Furthermore, the data were collected while subjects used passive prostheses. The introduction of an active prosthesis will likely alter the user's gait, potentially requiring iterative model updates.
Future research should focus on validating these models with larger cohorts and testing their robustness using noisier data from wearable sensors like IMUs. The ultimate goal is the real-time implementation of these Random Forest models as high-level controllers in physical prostheses to improve user comfort and metabolic efficiency.
Summary updated