Background: Falls risk in older people is multifactorial and complex. There is uncertainty about the importance of specific risk factors. Genetic algorithm neural networks (GANNs) can examine all available data and select the best nonlinear combination of variables for predicting falls. The aim of this work was to develop a risk profile for operationally defined new falls in a random sample of older people by use of a GANN approach.
Methods: A random sample of 1042 community-dwelling people aged 65 and older, living in Nottingham, England, were interviewed at baseline (1985) and survivors reinterviewed at a 4-year follow-up (n = 690). The at-risk group (n = 435) was defined as those survivors who had not fallen in the year before the baseline interview. A GANN was used to examine all available attributes and, from these, to select the best nonlinear combination of variables that predicted those people who fell 4 years later.
Results: The GANN selected a combination of 16 from a potential 253 variables and correctly predicted 35/114 new fallers (sensitivity = 31%; positive predictive value = 57%) and 295/321 nonfallers (specificity = 92%; negative predictive value = 79%); total correct = 76%. The variables selected by the GANN related to personal health, opportunity, and personal circumstances.
Conclusions: This study demonstrates the capacity of GANNs to examine all available data and then to identify the best 16 variables for predicting falls. The risk profile complements risk factors in the current literature identified by use of standard and conventional statistical methods. Additional data about environmental factors might enhance the sensitivity of the GANN approach and help identify those older people who are at risk of falling.