A machine learning approach to understanding falls in older women

Background: Falls are a significant public health concern for older adults, and are associated with many detrimental physical, psychological and economic consequences. The aim of this thesis was to identify which functional variables (i.e. balance, gait and clinical measures) and physical characteristics (i.e. strength and body composition) could best distinguish between older female fallers and non-fallers, using a novel machine learning approach.

Methods: 60 community-dwelling older women (≥65 years), classified as fallers and non-fallers based on falls history, attended three data collection sessions. Centre of pressure data were collected during five static posturography protocols. Ground reaction force (GRF) and kinematic data were measured during walking trials at two gait speeds. Participants completed five clinical measures, and strength assessments were performed at the trunk, knee and ankle for both limbs using an isokinetic dynamometer. Dual energy X-ray absorptiometry scans were conducted to assess body composition, bone mineral density and hip structure. The data (281 variables) were partitioned into five data packages (balance, gait, clinical measures, strength and body composition) and a uniform analysis strategy was applied to each in the single-domain analyses. Random forest and leave-one-variable-out partial least squares correlation analysis were employed to assess variable importance. The important variables selected were combined into two refined datasets, and classification models were constructed to differentiate between fallers and non-fallers. Subsequently, the strongest discriminators from each data package were compiled and analysed in the multi-domain analyses using the same strategy.

Findings: Overall, this thesis demonstrates that it is possible to distinguish between fallers and non-fallers with a high degree of accuracy, using a refined set of variables and a sophisticated multivariate approach, in both a single- and multi-domain context. Important balance, gait, clinical measures, strength, and body composition variables were identified for discriminating between groups, alongside redundant factors. In the balance data package, multiple variables measured during the limits of stability, unilateral stance, and Modified Clinical Test of Sensory Interaction on Balance were identified as important discriminators. In the gait data package, a combination of spatiotemporal, kinematic, GRF, and variability variables were selected as important. In the clinical measures data package, several variables measured during the Timed Up and Go, Tinetti Performance Oriented Mobility Assessment, and gait speed protocols appeared useful when differentiating between groups. In the strength data package, a combination of peak torque variables, namely knee flexion, dorsiflexion and plantar flexion, alongside three asymmetry variables were identified as important. In the body composition data package, several muscle quality indices were selected as important. In the multi-domain data package, a combination of gait, clinical measures, strength, and body composition variables appeared to be the most important discriminators between fallers and non-fallers. This study is one of the few comprehensive analyses to include important variables from multiple domains with a sophisticated machine learning approach. The findings can be used to inform the design of optimal falls screening and prevention methods. Furthermore, this work highlights the applicability of machine learning techniques when investigating falls in older women.