Age-Related Differences in Healthy Adults Walking Patterns Under a Cognitive Task With Deep Neural Networks

This 2020 IEEE Sensors Journal study uses a smart carpet and deep neural networks to examine how age shapes walking under a cognitive task in healthy adults.

Published in the IEEE Sensors Journal in September 2020, the paper by Omar Costilla-Reyes, Patricia Scully, Iracema Leroi, and Krikor B. Ozanyan grew out of the healthcare theme of Dr. Costilla Reyes’s doctoral research at the University of Manchester. There it appeared as the manuscript “Thinking and walking: deep neural networks to detect age-related differences in dual-tasks from healthy adults.” The same four collaborators held a 2018 Data Science Institute grant on gait patterns for early detection of cognitive decline.

A dual task asks a person to walk while also completing a cognitive activity. The extra mental load can change how people walk, and the size of the change may differ with age. The study used the UoM-Gait-69 dataset, recorded from healthy adults of different ages on the university’s plastic-optical-fiber smart carpet, a floor sensor that images footsteps by tomography, under an ethics protocol for dual-task analysis in healthy adults. Participants completed a series of experimental tasks. Raw sensor signals were split, cleaned, downsampled, smoothed, and arranged into spatio-temporal raw sensor matrices, which a deep neural network learned from directly, without hand-designed features. The network was trained to separate the dual-task walking conditions.

The deep model was compared with genetic programming, which evolves candidate programs, and with ensemble methods, which combine several models, and the learned features were visualized with t-SNE, a technique that maps high-dimensional features onto a plane so that clusters can be inspected. The deep network separated the dual-task conditions more accurately than the classical methods, and its best results were for the youngest group of walkers. The result shows that deep networks can detect the effect of a cognitive task on walking directly from raw floor-sensor signals, which gives researchers a protocol, a dataset, and a learning pipeline for studying how age shapes walking under mental load in healthy adults. It extends the pilot presented at IEEE SENSORS 2017, “Age-sensitive differences in single and dual walking tasks from footprint floor sensor data,” and draws on the thesis research described under “Pattern recognition from raw spatio-temporal data for gait analysis in healthcare and security.”