A Data Driven Approach for Predicting Preferred Ankle Stiffness of a Quasi-Passive Prosthesis

Emerging variable-stiffness ankle prostheses can modulate their stiffness to meet differing biomechanical demands. To this end, knowledge of the optimal ankle stiffness is required for each user and activity. One approach is to match the stiffness of prosthesis to the user's preference, but this requires a tuning process to determine each user's preferences. In this work, we seek to estimate user-preferred ankle stiffness using biomechanical data collected from seven subjects during walking at stiffness settings around their preferred stiffness; our hope is an automated method may reduce the time and experimental burden of determining user preferences. We investigated different machine learning algorithms, sensor subsets, and the impact of user-specific training data on estimation accuracy. We found that a long short term memory (LSTM) algorithm trained on user-specific data from only the affected side, were able to predict user preferred ankle stiffness with an RMSE of 5.2% \pm 0.3%. The prediction error was less than prosthesis users' ability to reliably sense stiffness changes (7.7%), which highlights the significance of the performance of our proposed method. This study provides a foundation for an automated approach for predicting user-preferred prosthesis mechanics that would ease the burden of tuning these systems in a clinical setting.