Developing a Falls Prediction Tool Using Both Accelerometer and Video Gait Analysis Data in Older Adults

Recent advances in machine learning have disrupted the standard approach to assessing medical prognosis. Our group, consisting of academic clinicians and research engineers, seeks to create a database of stability measures ( accelerometers, gyroscopes and altitude sensors data) in older adults monitores longitudinally. This Stability Measures (SM) database will allow us to use new machine learning methods to develop and validate algorithms that predict future falls, allowing to better targetting of vulnerable individuals.

Although there have been numerous attempts to quantify fall risk in older adults using bedside scales 7-12, no previous group has attempted to use a combination of both accelerometer and video measures to assess gait stability. Since these measures will be captured in both frequently falling and infrequently falling patients, we will have SM data for various windows of time (1, 2, 3 and 4-weeks) prior to at least 100 fall events, a dataset that has never been captured before.

HYPOTHESES:

1. A combination of accelerometer, gyroscope, and video data can be used to predict falls longitudinally, first by the use of a training dataset followed by verification on a validation data set.
2. All the above sensor-based inputs can be combined as a simple, automated predcition tool to predict fall risk in older adults Current Methods of Falls Risk Assessment: Current methods of predicting falls in physician offices rely heavily on simple bedside tests12-14. Although useful, all of these measures have quite low sensitivity and specificity, with an Area Under the Curve (AUC) of approximately 0.707-12.

In fact, a recent meta-analysis "could not identify any tool which had an optimal balance between sensitivity and specificity, or which was clearly better than a simple clinical judgment of risk of falling"

METHODS:

1. Subjects: i) High Risk Subjects (n=50): All subjects will be recruited from falls and geriatrics clinics at Vancouver General Hospital. These clinics see about 2500 patients per year and are currently used for research recruitment. Each clinic patient has gait speed measured, which will allow to recruit both high and low risk fallers. This test will allow us to recruit 50 subjects at marked risk for falls, providing us with prospectively gathered dataset of greater than 100 events, five times higher than any other sensor study. ii) Low Risk Subjects (n=50): In addition, we will use newspaper advertisements to recruit and then screen low risk subjects. All subjects will have a gait speed > 0.8 m/s and have had no falls in the last year.

All study patients with come to the laboratory (Gerontology and Diabetes Research Laboratory, VGH Research Pavilion) for a one hour session. Each subject will perform a 6-minute walk test during which gait assessment will be obtained from the APDM system (Portland, OR). In addition there will be four video cameras (on the front, back and sides) that will measure raw video data for our gait analysis. The camera does not record any facial data (in fact, 'deepfake' software in the system deletes all facial details) and the patient's movements are converted to a 'stick figure' prior to being saved in the system. In addition, a Xethru X4M03 kit was will be used to collect ultra-sideband radar data (UWB). The UWB radar operates in 5.9-10.3 GHz, providing high spatial resolution. The radar is placed 1.5 m above the floor level. To collect heel-toe strike timing data, the subject will ambulate on the GAITRite system (CIR Systems Inc, Franklin, NJ), with a 90 × 700-cm × 3.2-mm walkway.