Machine-Learning Based Determination of Gait Events from Foot-Mounted Inertial Units

Abstract

A promising but still scarcely explored strategy for the estimation of gait parameters based on inertial sensors involves the adoption of machine learning techniques. However, existing approaches are reliable only for specific conditions, inertial measurements unit (IMU) placement on the body, protocols, or when combined with additional devices. In this paper, we tested an alternative gait-events estimation approach which is fully data-driven and does not rely on a priori models or assumptions. High-frequency (512 Hz) data from a commercial inertial unit were recorded during 500 steps performed by 40 healthy participants. Sensors' readings were synchronized with a reference ground reaction force system to determine initial/terminal contacts. Then, we extracted a set of features from windowed data labeled according to the reference. Two gray-box approaches were evaluated: (1) classifiers (decision trees) returning the presence of a gait event in each time window and (2) a classifier discriminating between stance and swing phases. Both outputs were submitted to a deterministic algorithm correcting spurious clusters of predictions. The stance vs. swing approach estimated the stride time duration with an average error lower than 20 ms and confidence bounds between ±50 ms. These figures are suitable to detect clinically meaningful differences across different populations.

Keywords: decision trees; gait analysis; spatio-temporal parameters; wearable sensors.

Figure 1

GaitUp Physilog size (top) and placement on the foot (bottom). Direction of the local right-handed reference frame axes is also reported.

Figure 2

Experimental setup—tests were conducted on the middle laboratory lane with force platform embedded on the floor. Motion capture cameras are fixed on the wall in a standard gait analysis configuration.

Figure 3

Sample vertical (z-axis) raw accelerometer readings during a test. The first three spikes correspond to the synchronization signal and the following data refer to gait events.

Figure 4

Data processing flow: sensors’ readings (in the top panel, sample acceleration signal) were windowed (dashed blue boxes represent the window moving across the signal); subsequently a set of features were obtained for each window, which was labelled according to the reference ground reaction force output. The whole set of collected strides (each one containing a collection of features) were randomly split into a training and a test set. HS: heel-strike, TO: toe-off.

Figure 5

Correction algorithm: isolated short (counter < threshold) clusters were corrected according to the surroundings, as schematically reported in the bottom diagram.

Figure 6

Classification performance of the three tested approaches. Left: heel-strike (1a) and toe-off (1b) events detection; Right: stance vs. swing moving windows classification (2).

Figure 7

Regression plots comparing estimated and reference stride time (a) and the measurement error as a function of gait speed (b). Regression lines (dashed) and 95% confidence bounds (solid lines) were reported.