Clinical evaluation of a body-worn sensor-based system for fall risk testing

Abstract

Falls in older people cause morbidity and mortality, and are recognised as a unique geriatric syndrome. Despite major advances and concentrated efforts, fall rates have not decreased in recent decades, leading clinicians and researchers to search for effective fall prevention approaches and better mechanisms for allocating limited interventional resources. This work focuses on screening and assessment of fall risk in older individuals. Specifically, it involves the use of a miniaturised, waist-worn accelerometer for Sensor-based Fall Risk Testing (SFRT).

Previous work by our research team involved older community-dwellers who performed a certain movement routine while wearing the sensor device; they also underwent a known fall risk assessment. Features extracted from the sensor signals were used to map against the clinical assessment scores. The work in this thesis presents a continuation of this earlier study, initially by testing the feasibility of unsupervised conduct of the routine by older individuals. The results of a pilot study, simulating an unsupervised sensor-based assessment, revealed substantial concerns with regards to both safety and correct, unassisted execution of the said routine.

The routine was then expanded and incorporated into a fall prevention program at a public clinic serving older community-dwellers. The results of this larger, prospective study demonstrated that the use of SFRT is in fact feasible in such settings, but raised notable clinical issues which have yet to be addressed in the pertinent literature. The study also demonstrated the potential of this approach in evaluating the efficacy of the exercise-based intervention. Finally, features extracted from the sensor signals were incorporated into negative binomial regression models. Fall counts, collected over a follow-up period of up to a year following the assessment, were used as the target. A deliberate effort was made to avert common methodological pitfalls that result in model over-fitting and inappropriate model validation. An attempt was made to train and validate models using prospective falls data from the aforementioned earlier study. Significant differences between the participant cohorts largely resulted in model rejection during the validation phase, stressing the critical importance of proper external validation in developing clinically-relevant and robust models that possess true predictive power.