Enhancing neuroprosthesis calibration: the advantage of integrating prior training over exclusive use of new data.

Abstract

Neuroprostheses typically operate under supervised learning, in which a machine-learning algorithm is trained to correlate neural or myoelectric activity with an individual's motor intent. Due to the stochastic nature of neuromyoelectric signals, algorithm performance decays over time. This decay is accelerated when attempting to regress proportional control of multiple joints in parallel, compared with the more typical classification-based pattern recognition control. To overcome this degradation, neuroprostheses and commercial myoelectric prostheses are often recalibrated and retrained frequently so that only the most recent, up-to-date data influences the algorithm performance. Here, we introduce and validate an alternative training paradigm in which training data from past calibrations is aggregated and reused in future calibrations for regression control.
Approach: Using a cohort of four transradial amputees implanted with intramuscular electromyographic recording leads, we demonstrate that aggregating prior datasets improves prosthetic regression-based control in offline analyses and an online human-in-the-loop task. In offline analyses, we compared the performance of a convolutional neural network (CNN) and a modified Kalman filter (MKF) to simultaneously regress the kinematics of an eight-degree-of-freedom prosthesis. Both algorithms were trained under the traditional paradigm using a single dataset, as well as under the new paradigm using aggregated datasets from the past five or ten trainings. 
Main Results: Dataset aggregation reduced the root-mean-squared error of algorithm estimates for both the CNN and MKF, although the CNN saw a greater reduction in error. Further offline analyses revealed that dataset aggregation improved CNN robustness when reusing the same algorithm on subsequent test days, as indicated by a smaller increase in RMSE per day. Finally, data from an online virtual-target-touching task with one amputee showed significantly better real-time prosthetic control when using aggregated training data from just two prior datasets. 
Significance: Altogether, these results demonstrate that training data from past calibrations should not be discarded but, rather, should be reused in an aggregated training dataset such that the increased amount and diversity of data improve algorithm performance. More broadly, this work supports a paradigm shift for the field of neuroprostheses away from daily data recalibration for linear classification models and towards daily data aggregation for non-linear regression models.