Turn Intent Detection For Control of a Lower Limb Prosthesis

Objective: An adaptable lower limb prosthesis with variable stiffness in the transverse plane requires a control method to effect changes in real time during amputee turning. This study aimed to identify classification algorithms that can accurately predict turning using inertial measurement unit (IMU) signals from the shank with adequate time to enact a change in stiffness during the swing phase of gait when the prosthesis is unloaded. Methods: To identify if a turning step is imminent, classification models were developed around activities of daily living including 90° spin turns, 90° step turns, 180° turns, and straight walking using simulated IMU data from the prosthesis shank. Three classifiers were tested: support vector machine (SVM), K nearest neighbors (KNN), and a bagged decision tree ensemble (Ensemble). Results: Individual training gave superior results over training on a pooled set of users. Coupled with a simple control scheme, the SVM, KNN, and Ensemble classifiers achieved 96%, 93%, and 91% accuracy (no significant difference), respectively, predicting an upcoming turn 400 ± 70 ms prior to the heel strike of the turn. However, classification of straight walking transition steps varied between classifiers at 85%, 82%, 97% (Ensemble significantly different, p = 0.002), respectively. Conclusion: The Ensemble model produced the best result overall; however, depending on the priority of identifying turning versus transition steps and processor performance, the SVM or KNN might still be considered. Significance: This research would be useful to help determine a classifier strategy for any lower limb device seeking to predict turn intent.