Online Learning of Gait Models

Abstract

Gait event identification is the identification of gait events (e.g., foot impact) which occur cyclically, at the same location in each gait cycle. Gait event identification plays an important role in many applications, such as health monitoring, diagnosis, and rehabilitation. The majority of gait event identification algorithms are based on heuristics, many of which are threshold-based, making them sensitive to threshold parameters and causing poor generalization to new data (e.g., different gait type, ground surface, sensor placement, footwear, etc.). While a number of machine learning techniques have been proposed, they use offline training and do not generalize well to data that is different from the training set.

This thesis proposes a novel approach for online, individualized gait analysis, based on an adaptive periodic model of any gait signal. The proposed method learns a model of the gait cycle during online measurement, using a continuous representation that can adapt to inter and intra-personal variability by creating an individualized model. The model of gait is learned online during observation, using incremental updates to the model parameters based on the error between the model-predicted and measured signal. The gait data is modeled as a periodic signal with a continuous phase variable, allowing data to be automatically labeled with a phase value corresponding to a particular event that re-occurs each gait cycle. Once the algorithm has converged to the input signal, key gait events can be identified based on the estimated gait phase.

Two methods of gait event identification were implemented: analytical event identification and initial event identification. Since we learn a gait model that has an analytical representation, if we know the properties of the gait events of interest, we can use the model to directly compute the corresponding phase. For example, to identify the peak event in each gait cycle, we can solve for the phase which generates the maximum value and assign this as the peak phase value. In the initial event identification method, we provide a manual or heuristic identification of a gait event in the first converged gait cycle and use the corresponding event phase to identify all future events. Once gait events are identified relative to the estimated gait phase, we can automatically identify any future events since we assume they occur at the same phase, as is common in gait analysis.

Our approach is implemented and tested on two datasets: one measuring medio lateral angular velocity of the ankles from a healthy young group of adults and the other measuring sagittal linear acceleration of the ankles from a group of retirement home residents who each have a variety of medical conditions. For the former dataset, the proposed approach converges within approximately five gait cycles and heel impact and toe takeoff events are extracted with an average error of 0.04 gait cycles, using the manual initial event identification method.

For the latter dataset, the proposed approach converges within approximately eight gait cycles and initial swing events are extracted with an average error of 0.03 gait cycles, using the analytical event identification method. When using learning rates optimized on a set of training trials (opposed to a default set of learning rates), the proposed approach converges within approximately four gait cycles and maintains an average error of 0.03 gait cycles, on the corresponding set of test trials. Further, when including ground truth events occurring prior to the model having met convergence criteria, the average error is only slightly increased to 0.04 gait cycles.