AbstractsBiology & Animal Science

Falls are a prevalent and significant problem facing a growing older population. Many existing clinical fall risk assessment tools are complicated, time-consuming, and involve subjective input from either clinical practitioners or the older individuals being assessed. In addition, most clinical assessment tools have undergone only limited independent validation to ascertain their true accuracy in predicting future falls. Body-worn sensor-based movement analysis shows great promise as a means of assessing fall risk, given that it enables quantitative signal analysis and can be integrated into a simple assessment procedure. Such body-worn sensor-based systems could also feasibly be used for unsupervised assessment and long-term monitoring. Previous studies within the author's research group have attempted to construct a waist-worn accelerometry-based system, using time domain features extracted from the accelerometry signals, which were acquired from the waist-worn triaxial accelerometer (TA) as the study participant executed a scripted movement routine. These studies aimed to develop a model which could estimate fall risk for the 68 elderly subjects, by mapping to a clinical fall risk assessment score. Based on this work, this thesis report firstly presents an attempt to improve this model, using spectral analysis of the same accelerometry signals. However, a strong optimistic bias was discovered in how validation (estimating future model performance) was previously performed for this modelling technique; the modelling method is adjusted to remove this validation bias. Another major challenge encountered relates to the large feature set dimensionality and small available sample size, hence a dimensionality reduction method is employed prior to model training. Subsequently, as one of its major contributions, this thesis provides an external validation of this waist-worn sensor-based fall risk assessment system using an independent group of subjects recruited into a validation study. However, the results of this validation show poor performance in estimating the reference clinical fall risk test scores which the proposed sensor-based fall risk assessment model attempts to estimate. This thesis project also presents a Poisson regression model to estimate the fall rates of the subjects over a period of several months (up to 13 months) following an initial sensor-based assessment, using the same group of accelerometry-based features from the proposed directed movement routine. An independent validation is also performed using assessment data from the same validation study; however, the validation performance is also poor. A discussion of the necessity and disadvantages of the dimensionality reduction process used prior to modelling is raised, highlighting how some useful features are discarded during this procedure, but that this process is entirely necessary in order to mitigate the risk of model over-fitting. Subsequently, a TA-based reaction time (RT) test and a new walking task (the six meter walk test…