Predictive Models For Falls-Risk Assessment in Older People, Using Markerless Motion Capture

Maudsley-Barton, S (2020) Predictive Models For Falls-Risk Assessment in Older People, Using Markerless Motion Capture. Doctoral thesis (PhD), Manchester Metropolitan University.

Preview

Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (7MB) | Preview

Abstract

Falling in old age contributes to considerable misery for many people. Currently, there is a lack of practical, low cost and objective methods for identifying those at risk of falls. This thesis aims to address this need. The majority of the literature related to falls risk and balance impairment uses force plates to quantify postural sway. The use of such devices in a clinical setting is rare, mainly due to cost. However, some force-plate-based commercial products have been created, e.g. the Balance Master. To align the research in this thesis to both the literature and existing methods of assessing postural sway, a method is proposed which can generate sway metrics from the output of a low-cost markerless motion capture device (Kinect V2). Good agreement was found between the proposed method and the output of the Balance Master. A key reason for the lack of research into falls-risk using markerless motion capture, is the lack of an appropriate dataset. To address this issue, a dataset of clinical movements, recorded using markerless motion capture, was created. Named KINECAL, It contains the recordings of 90 participants, labelled by age and falls-risk. The data provided includes depth images, 3D joint positions, sway metrics and socioeconomic and health meta data. Many studies have noted that postural sway increases with age and conflate age-related changes with falls risk. However, if one examines sub-populations of older people, such as master athletes, It is clear that this is not necessarily true. The structure of KINECAL allows for the examination of age-related factors and falls-risk factors simultaneously. In addition, it includes labels of falls history, clinical impairment and comprehensive metadata. KINECAL was used to identify sway metrics most closely associated with falls risk, as distinct from the ageing process. Using the identified metrics, a model was developed that can identify those who would be classified as impaired by a range of clinical tests. Finally, a model is proposed, which can predict fallers by placing individuals on a scale of physical impairment. An autoencoder was used to model, healthy adult sit-to-stand movements. Using an anomaly detection approach, an individuals level of impairment can be plotted relative to this healthy standard. Using this model, the existence of two older populations (one with a high falls risk and one with a low falls risk) is demonstrated.