Detecting Disorientation of People with Dementia

Real-time detection of spatial disorientation from wearable sensor data is a highly relevant task in the context of situation-aware, proactive navigation assistance technology. Such a detection system can be based on supervised learning of a machine learning model that discriminates normal and disorientation states. Ideally, the training (and test) data would be obtained in a field study consisting of a real-world orientation task, but such field studies can be challenging, and are not guaranteed to cover the disorientation instances one is interested in, due to the large variability of the environment.

We developed an experimental setup that allows to obtain motion trajectories of human protagonist while performing a navigation task in a lab environment. This is done by posing the navigation task in a virtual reality (VR) environment projected in front of a treadmill-based gait analysis system, such that walking on the treadmill results in movement in the VR environment. This setup allows to induce disorientation, by manipulating landmarks in the environment. The results of the study show that gait and accelerometric features are strongly associated with disorientation, thus showing a promising direction for automatically detecting spatial disorientation from gait data.