Background
A key factor in successfully preventing falls, is early identification of elderly with a high risk of falling. However, currently there is no easy-to-use pre-screening tool available; current tools are either not discriminative, time-consuming and/or costly. This pilot investigates the feasibility of developing an automatic gait-screening method by using a low-cost optical sensor and machinelearning algorithms to automatically detect features and classify gait patterns.

Method
Participants (n = 204, age 27 ± 7 yrs.) performed a gait test under two conditions: control and with distorted depth perception (induced by wearing special goggles). Each test consisted of 4x 3m walking at comfortable speed. Full-body 3D kinematics were captured using an optical sensor (Microsoft Xbox One Kinect). Tests were conducted in a public space to establish relatively 'natural' conditions. Data was processed in Matlab and common spatiotemporal variables were calculated per gait section. The 3D-time series data of the centre of mass for each section was used as input for a neural network, that was trained to discriminate between the two conditions.

Results
Wearing the goggles affected the gait pattern significantly: gait velocity and step length decreased, and lateral sway increased compared to the control condition. A 2-layer neural network could correctly classify 79% of the gait segments (i.e. with or without distorted vision).

Conclusions
The results show that gait patterns of healthy people with distorted vision could automatically be classified with the proposed approach. Future work will focus on adapting this model for identification of specific physical risk-factors in elderly.