A machine learning approach to identify important variables for distinguishing between fallers and non-fallers in older women.

Abstract

Falls are a significant ongoing public health concern for older adults. At present, few studies have concurrently explored the influence of multiple measures when seeking to determine which variables are most predictive of fall risks. As such, this cross-sectional study aimed to identify those functional variables (i.e. balance, gait and clinical measures) and physical characteristics (i.e. strength and body composition) that could best distinguish between older female fallers and non-fallers, using a machine learning approach. Overall, 60 community-dwelling older women (≥65 years), retrospectively classified as fallers (n = 21) or non-fallers (n = 39), attended three data collection sessions. Data (281 variables) collected from tests in five separate domains (balance, gait, clinical measures, strength and body composition) were analysed using random forest (RF) and leave-one-variable-out partial least squares correlation analysis (LOVO PLSCA) to assess variable importance. The strongest discriminators from each domain were then aggregated into a multi-domain dataset, and RF, LOVO PLSCA, and logistic regression models were constructed to identify the important variables in distinguishing between fallers and non-fallers. These models were used to classify participants as either fallers or non-fallers, with their performance evaluated using receiver operating characteristic (ROC) analysis. The study found that it is possible to classify fallers and non-fallers with a high degree of accuracy (e.g. logistic regression: sensitivity = 90%; specificity = 87%; AUC = 0.92; leave-one-out cross-validation accuracy = 63%) using a combination of 18 variables from four domains, with the gait and strength domains being particularly informative for screening programmes aimed at assessing falls risk.