The influence of sensor placement on the shoe on the performance of CNN based gait analysis algorithms

Gait parameters derived from foot-worn inertial measurement units (IMUs) have been investigated extensively as a digital mobility marker for a variety of diseases. However, it remains poorly understood how the performance of algorithms for spatial gait parameters change when the placement of the sensors is varied. Only few studies investigated this influence, and they only tested physics based algorithms. Over the past years, machine learning and in particular neural networks have been proposed as an alternative method to estimate spatial gait parameters. It is an open question, how consistent these new methods are when the sensor is placed at different positions. Hence, we studied the differences in performance of two convolutional neural network architectures for stride length estimation across six sensor positions on the feet. We conducted an empirical study on a publicly available dataset consisting of 8,989 strides from 14 participants with stride length reference based on motion capture. Our results indicate that the performance of both network architectures are sensitive to changes in the sensor position. Further, we found a consistent relationship between the different sensor positions and the stride length error for both approaches and the physics-based algorithms studied in previous work. These findings motivate further investigations into the origins of the observed differences and the robustness of gait analysis systems in general, to ensure more reliable measurements under unsupervised real-world conditions in the future.