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

Objective: This paper describes the developmentand preliminary offline validation of an algorithm facilitatingautomatic, self-contained learning of ground terrain transitionsin a lower limb prosthesis. This method allows for continuous,in-field convergence on an optimal terrain prediction accuracyfor a given walking condition, and is thus not limited bythe specific conditions and limited sample size of an in-labtraining scheme. Methods: We asked one subject with a below-kneeamputation to traverse level ground, stairs, and rampsusing a high-range-of-motion powered prosthesis while internalsensor data were remotely logged. We then used these datato develop a dynamic classification algorithm which predictsthe terrain of each stride and then continuously updates thepredictor using both data from the previous stride and anaccurate terrain back-estimation algorithm. Results: Across 100simulations randomizing stride order, our method attained amean next-stride prediction accuracy of ? 96%. This valuewas first reached after ? 200 strides, or about ? 5 minutesof walking. Conclusion and significance: These results demonstratea method for automatically learning the gait patternspreceding terrain transitions in a prosthesis without relyingon any external devices. By virtue of its dynamic learningscheme, application of this method in real-time would allow forcontinuous, in-field optimization of prediction accuracy across avariety of walking variables including physiological conditions,variable terrain geometries, control methodologies, and users.