Neural network-based prediction of missing key features in vertical GRF-time recordings

Ground reaction force (GRF) time recordings are frequently corrupted due to faulty stepping on a force platform during gait, and this results in the loss of valuable data and the need for additional data acquisition. This paper proposes a new approach based on artificial neural networks for the estimation of lost key gait parameters using the available features of an affected GRF-time trace. GRF-time plots were recorded using a force platform during normal walking of 14 young and 13 elderly individuals. Back-propagation neural network models were developed using features extracted from 466 vertical GRF-time characteristics that would be unaffected as inputs, and the likely affected gait features (e.g. stance time (ST), push-off force, (Fmax2) and push-off time (Tmax2)) as output. Performance of the models in predicting ST, Fmax2 and Tmax2 were tested using data from 30 new gait trials. The neural network model predicted the missing STs with 96.5% (+/-2.6%) accuracy (r>0.9). Accuracy of the ST prediction deteriorated when push-off force/time data were unavailable for the prediction model. Fmax2 and its time were reconstructed with an accuracy of 95.7% (+/-2.8%) and 97.6% (+/-1.7%) respectively. These results suggest that an artificial neural network may be applied to estimate missing ST, Fmax2 or Tmax2 information using features from an affected vertical GRF-time plot, and the method shows good promise for reconstructing gait forces from a corrupted force-time trace.