Elham Dolatabadi, Twinkle Arora, Andrea Iaboni, and Babak Taati
Falls are common and lethal in people with dementia. However, current fall risk assessments in dementia care units are static, non-specific, and lacking clinical utility. To improve on these assessments, we believe if we monitor and analyse all the risk factors for fall longitudinally over time, we can flag an individual who is unwell and at risk of falling. This study aims to advance the use of vision-based technology for unobtrusive and longitudinal monitoring of falls risk factors, and the development of a new paradigm of dynamic falls risk prediction in dementia.
We included longitudinal and baseline data from 45 patients at the Specialized Dementia unit at the Toronto Rehabilitation Institute – University Health Network. For each patient in the database, baseline data including demographics and clinical assessments as well as longitudinal data including gait episodes (over the course of the day and night), pharmacy records of medications, falls, and restraint use were gathered during their length of stay (from admission to discharge).
Deep learning model
We developed a novel deep learning model based on Recurrent Neural networks (RNN) called composite RNN. The composite RNN was built on wide range of longitudinal and baseline data to predict the risk of falls in short term (days). It is a novel approach which handles three common methodological challenges arising from clinical data: (a) it learns a rich representation of high dimensional sequences of gait episodes and medications; (b) it handles unequal time intervals between the recorded sequences; and (c) it combines a wide range of longitudinal, categorical, and baseline clinical data for prediction.
This model outperformed traditional machine learning predictive models in all cases. We believe that this approach can be used to create accurate detection of an individual who is unwell and at risk of falling.