Cognition in healthy older women is a predictor of 14‐year falls risk

Abstract

AbstractBackgroundFalls are a significant cause of injuries, loss of confidence, increased morbidity, and institutionalisation in all older people, with women at 50% greater risk than men. The relationship between dementia and falls is well established and 2/3 of all dementia occurs in women. In this study we explored risk factors associated with a 14 year falls risk in a community‐based cohort of women, which included validated measures across a wide range of clinical domains including neuropsychological, mood, quality of life and biomarkers (including hormonal).MethodThe Australian Women’s Healthy Aging Project is an longitudinal observation study, assessments every year (1991 –1999), followed by assessments in 2002, 2004, 2012 and 2014. The assessments included cognitive (as of 2002), blood, and cardiovascular disease risk assessment, and questions related to falls. After data cleaning, the remaining cohort consisted of 180 participants (Table 1). Missing data were imputed using mice random forest. To identify key risk factors associated with a 14 year falls risk, random survival (time to event) forest (RSF) machine learning was used.ResultThe RSF model, using all 290+ possible predictive variables, performed well with an Out Of Bag (OOB, withheld data) prediction error (C‐index) of 32.8%. The most predictive variables in the model were identified using the variable importance measure (VIM).The initial model was refined by taking the top 30 predictive variables and retraining the RSF. This refined model resulted in an improved OOB C‐index of 5.8% (27%). The top 20 predictive variables, Figure 1, include those associated with cardiovascular disease risk, cognitive performance, and hormone levels (e.g., family history of heart attack, digit symbol coding, and estradiol levels).ConclusionNinety percent of the top 20 predictive risk variables for the 14 year fall risk in women, were from three key domains, cognition (40%), cardiovascular (25%) and hormone‐related measurements (25%). Our data suggest that for long term prevention of falls these domains may be important reducing risk of falls in the senior female population.