Uncertainties of 3-D Motion Prediction From a Single Virtual Accelerometer During Walking and Running

Although studies have demonstrated the potential of using a single inertial sensor for walking and running motion prediction, further analysis of sensor position and training dataset size is required. This study investigates the accuracies and two types of uncertainties associated with using a single virtual accelerometer for whole-body motion predictions. Motion capture data during walking and running on a treadmill at four speeds were obtained from 198 participants. A Bayesian neural network (BNN) was trained to predict kinematic gait parameters from virtual accelerometer data. The root-mean-squared error (RMSE) ranged from 1.7° to 10.0° depending on three sensor positions– left wrist, lower back, and right shank. As the dataset size increased from 20 to 160 subjects, error and epistemic uncertainty decreased (
${p} < 0.001$
). However, aleatoric uncertainty remained constant (
${p}$
= 0.247) based on the linear mixed-effect model. We observed sensor positions that predicted each gait parameter with the lowest error and uncertainty. For example, the sensor on the lower back showed less right hip flexion prediction errors (RMSE
$4.04^{\circ }{)}$
than those on the left wrist (RMSE
$5.10^{\circ }{)}$
and right shank (RMSE
$4.29^{\circ }{)}$
. The two uncertainties corresponded to different sources of prediction error. The optimal sensor position was found for each target joint motion along with the uncertainties. We suggested a regression network that can quantify errors from two different sources, training dataset size, and system limitations such as sensor position. It can be used to improve data-driven motion prediction using wearable devices by understanding the limitations.