A Machine Learning Based Strategy for Predicting the Sagittal Angles in real-time for Actuating an Active Knee-Ankle Prosthesis

Objectives Transfemoral amputees use knee-ankle prostheses to restore impaired locomotion. Active prostheses can potentially overcome the limitations of passive ones by emulating a more biological behavior, thus supporting a more natural gait. The control schemes for actuating the joints of such active prosthesis should be tailored to understand the user’s locomotive intention and translate them into the required locomotion similar to that of an able-bodied individual. Thus, the control schemes should be able to suit a real-time scenario and be fast enough for updating the model parameters online and predict the control commands in real-time. For individuals with transfemoral amputation, the control scheme should produce the desired locomotion by actuating an active knee and ankle joint to generate appropriate knee and ankle angles. Machine learning models could be potentially utilized to develop control schemes for such active prostheses.

Methods In this study, we employed a random forest (RF) regression model to learn a non-linear relationship between the thigh angular kinematics and the sagittal angles of the knee and ankle during walking. Once the RF model was trained, it was used to continuously predict the angles required at the knee and ankle during walking using thigh angular motion data as the inputs. RF is a machine learning model, which makes predictions by averaging the predictions of many decision estimators in the ensemble. RF regression was chosen as the model because of several advantages that make it suitable for real-time prediction. First, feature normalization is not necessary. Second, it takes less training data to reach a stable prediction accuracy and less training time compared to many other machine learning algorithms like Gaussian process regression and complex deep learning models. Third, it can handle a large input dimension without overfitting.

To train an RF model, we used motion capture data from an able-bodied subject, walking at self-selected speed in a gait laboratory. The RF model was constructed with 200 decision estimators pruned to a depth of six to enhance generalization ability to varying gait patterns during walking. The model was trained and validated with a leave-k-out cross-validation strategy (k=6) over 16 walking trials.

Results and Conclusion For level ground walking at self-selected speed, the proposed method could predict the angles of the knee and ankle with high accuracy (mean R-squared value of 0.995 for knee angle, 0.966 for ankle angle). The training time for the algorithm was 800 milliseconds for 10 trials of 250 samples each, and the prediction time was around 12 milliseconds for 250 samples.

The proposed strategy shows potential for continuously controlling an active ankle-knee prosthesis for a transfemoral amputee, in real-time, whose thigh angular motion can be used to infer the required prosthetic moments and angles.

Even without normalization of features, very good to excellent prediction results were obtained.

Stichwörter -

Publication History

Article published online:
15 October 2020