Gait event detection using a multilayer neural network

Manual detection of gait events via visual inspection of motion capture data is a laborious process. There are currently no robust techniques available to automate the process for pathologic gait. However, the detection of gait events is essentially a classification problem; an application for which artificial neural networks are well suited. In this paper, a multilayer artificial neural network is presented for the purpose of classifying foot-contact and foot-off events using the sagittal plane coordinates of heel and toe markers. The timing of events detected using this method was compared to the timing of events detected by measuring the ground reaction force using a force plate for a total of 40 pathologic subjects divided into two groups: barefoot and shod/braced. On average, the neural network detected foot-contact events 7.1 ms and 0.8 ms earlier than the force plate for the barefoot and shod/braced groups respectively. The average difference for foot-off events was 8.8 ms and 3.3 ms. Given that motion capture data were collected at 120 Hz, this implies that the force plate method and neural network method generally agreed within 1–2 frames of data. Consequently, the neural network was shown to be an accurate, autonomous method for detecting gait events in pathologic gait.