Preliminary Investigation Of Predicting Time-To-Next Heelstrike Using Accelerometers And Machine Learning

Osteoarthritis knee braces require large brace-leg interface forces to stabilize and unload the joint during weight bearing. Actively removing support while the user is in a non-weight-bearing state could improve the comfort of the brace but requires the timing of weight-bearing states to be known. This study presents two artificial neural networks (ANNs) for predicting time-to-next heelstrike during walking using only data from two accelerometers placed on the thigh and shank. One ANN used teacher forcing and the other did not. Walking data were collected from 10 subjects and leave-one-subject-out cross-validation was used to evaluate the performance of the two models. Input features for the ANNs included tibial and femoral accelerations, concatenated into one array. The teacher forcing ANN and the non-teacher forcing ANN performed equally well (RMSE = 0.23 +/- 0.13s for the non-teacher forcing ANN, RMSE = 0.27 +/- 0.08s for the teacher forcing ANN). The performances of the models were worse than those of previously published studies that predicted heelstrike events. Accelerations were insufficient for an ANN to predict time-to-next heelstrike during walking.