Search

Falls in older adults are a major problem; the current fall risk assessments used in fall prevention strategies suffer from a range of limitations. The use of wearable inertial sensors (IMUs) to quantitatively assess gait and predict falls can help fill these gaps. The aims of this research were focused on exploring the utility of sensor-based gait analysis on different terrains for predicting falls, which has so far received limited attention. Activity trials were conducted with two cohorts of community-dwelling older adults (N1=81, N2=62), who walked under semi-free conditions on flat surfaces and up and down stairs. They completed the clinically-validated Physiological Profile Assessment (PPA) and prospective falls were recorded over 12 months of follow-up. IMUs were worn on their lower back and right ankle; gait parameters were derived from the inertial data in two ways. In the first method, step rate, variability and vigour were derived from the sensors’ local-frame-of-reference. The second method calculated stride trajectories from the ankle sensor in the 3D global-frame-of-reference. The trajectory derivation algorithm was validated using motion capture data from healthy adults (N=19) walking on all three terrains at various speeds. From the trajectories, a selection of clearance and timing-related parameters were derived. Differences between terrains were quantified; correlations were tested between parameters and PPA components, as well as the incidence of multiple falls, for both frames of reference. The two cohorts, one for model-training and the other for validation, were analysed separately. In cohort 1, people who descended stairs with a faster step rate, shorter stride time, and lower ankle stride clearance were significantly more likely to experience multiple falls during follow-up. In cohort 2, these relationships were not found. Logistic regression models based on gait parameters were built and validated independently, by splitting the participants into separate training or validation datasets prior to modelling. The models achieved moderate classification accuracy overall (52-87%), with high specificity (55-100%) but lower sensitivity (0-44%), highlighting potential issues in the development of robust and powerful fall prediction models with different cohorts and limited sample sizes. This research demonstrated the potential value of stair-gait analysis in falls research.