Automatic Incremental Learning of Terrain Transitions in a Powered Below Knee Prosthesis

Abstract

Objective: This paper describes the development and preliminary offline validation of an algorithm facilitating automatic, self-contained learning of ground terrain transitions in a lower limb prosthesis. This method allows for continuous, in-field convergence on an optimal terrain prediction accuracy for a given walking condition, and is thus not limited by the specific conditions and limited sample size of an in-lab training scheme. Methods: We asked one subject with a below-knee amputation to traverse level ground, stairs, and ramps using a high-range-of-motion powered prosthesis while internal sensor data were remotely logged. We then used these data to develop a dynamic classification algorithm which predicts the terrain of each stride and then continuously updates the predictor using both data from the previous stride and an accurate terrain back-estimation algorithm. Results: Across 100 simulations randomizing stride order, our method attained a mean next-stride prediction accuracy of ? 96%. This value was first reached after ? 200 strides, or about ? 5 minutes of walking. Conclusion and significance: These results demonstrate a method for automatically learning the gait patterns preceding terrain transitions in a prosthesis without relying on any external devices. By virtue of its dynamic learning scheme, application of this method in real-time would allow for continuous, in-field optimization of prediction accuracy across a variety of walking variables including physiological conditions, variable terrain geometries, control methodologies, and users.