Development of a Convolutional Neural Network to estimate the Margin of Stability starting from IMU data

Gait stability indices quantify the risk of falling and can provide an early warning. Margin of Stability (MoS) is one of the most used indices and currently, the gold standard to measure it is motion capture systems (MoCap) that, however, presents some drawbacks. To overcome these limits, we proposed a Convolutional Neural Network (CNN) algorithm to estimate MoS values starting from IMUs data. This project comes from a collaboration with the research group of Prof. Akiyama at Shinshu University (Ueda, Nagano, Japan) and presents three aims: to create a CNN able to perform regression task; to compare the CNN outcomes while inserting in input gyroscope images versus accelerometer ones; to perform a GradCAM analysis, to find the more meaningful images areas for the CNN and then, to isolate a subgroup of the most interesting sensors and observe the performances. To fulfill these goals, two groups of participants were acquired with both MoCap and IMU sensors. The “Heterogeneous dataset” (HeD): 9 subjects who walked on a treadmill at two different speeds and while wearing or not a leg orthosis. The “Homogeneous dataset” (HoD): 10 people who performed level ground walking at comfortable speed. After images creation from the 2 datasets, we input them in the CNN which performed a regression task to estimate the MoS. The CNN outcomes did not show significant differences using accelerometer images with respect to gyroscope ones. From the HoD we obtained images with a maximum dimension [8 101], while from the HeD [6 101]. Then, after the CradCAM analysis we found out that the meaningful sensors for the CCN are the ones placed on the two feet for HeD and on L5 for HoD, so we tested images [2 101] and [1 101]. We conclude that there is no difference in using 8, 6, 2 (or 1) sensors and so a lower number IMUs can be used in future. As a future development, additional subjects can be added to enrich the actual datasets and evaluate possible improvements in CNN performance.