Real-Time Adaptation of an Artificial Neural Network for Transfemoral Amputees Using a Powered Prosthesis

Abstract

Objective: We evaluated a two-step method to improve control accuracy for a powered prosthetic leg using machine learning and adaptation, while reducing subject training time. Methods: First, information from three transfemoral amputees was grouped together, to create a baseline control system that was subsequently tested using data from a fourth subject (user-independent classification). Second, online adaptation was investigated, whereby the fourth subject’s data were used to improve the baseline control system in real-time. Results were compared for user-independent classification and for user-dependent classification (data collected from and tested in the same subject), with and without adaptation. Results: The combination of a user-independent classifier with real-time adaptation based on a unique individual’s data produced a classification error rate as low as 1.61% [0.15 standard error of the mean (SEM)] without requiring collection of initial training data from that individual. Training/testing using a subject’s own data (user-dependent classification), combined with adaptation, resulted in a classification error rate of 0.9% [0.12 SEM], which was not significantly different (P > 0.05) but required, on average, an additional 7.52 hours [0.92 SEM] of training sessions. Conclusion and Significance: We found that the combination of a user-independent dataset with adaptation resulted in error rates that were not significantly different from using a user-dependent dataset. Furthermore, this method eliminated the need for individual training sessions, saving many hours of data collection time.

A-Mode Ultrasound-Based Prediction of Transfemoral Amputee Prosthesis Walking Kinematics via an Artificial Neural Network

“…Although most controllers rely solely on mechanical sensors [9][10], camera systems have also been developed in an effort to improve the reliability of the user's intention detection [11] [12][13] [14]. Unfortunately, these techniques require training on large datasets, which can be resource intensive and scale poorly because subject specific datasets are typically necessary to improve accuracy of the classification [15], although real-time adaptation models may in part address this issue [10]. These classifiers are typically used to identify the intended ambulation mode and environment.…”

Up to one's knees in data : data-driven intent recognition using electromyography for the lower limb

“…We used a mode-specific Neural Network classifier, comparable as described by Woodward et al [125,126]. The neural network contained one hidden layer with 20 nodes as proposed by Woodward et al [125] and relu activation.…”

“…• Sample selection based on confidence using entropy [120,122,134] • Sample labelling based on backward prediction [123,124,126] • Combination of backward prediction and updating using entropy…”

“…a resampling procedure based on clustering [117][118][119], sample selection based on entropy [120][121][122] or sample selection based on backward prediction [123][124][125][126][127][128].…”

“…Woodward et al [126] used the backward predictor as well, using a feed forward Neural Network for the forward and backward predictors. They measured four transfemoral amputees over four sessions.…”