Falling in the elderly: Do statistical models matter for performance criteria of fall prediction? Results from two large population-based studies.

Objective: To compare performance criteria (i.e., sensitivity, specificity, positive predictive value, negative predictive value, area under receiver operating characteristic curve and accuracy) of linear and non-linear statistical models for fall risk in older community-dwellers.

Methods: Participants were recruited in two large population-based studies, "Prévention des Chutes, Réseau 4" (PCR4, n=1760, cross-sectional design, retrospective collection of falls) and "Prévention des Chutes Personnes Agées" (PCPA, n=1765, cohort design, prospective collection of falls). Six linear statistical models (i.e., logistic regression, discriminant analysis, Bayes network algorithm, decision tree, random forest, boosted trees), three non-linear statistical models corresponding to artificial neural networks (multilayer perceptron, genetic algorithm and neuroevolution of augmenting topologies [NEAT]) and the adaptive neuro fuzzy interference system (ANFIS) were used. Falls ≥1 characterizing fallers and falls ≥2 characterizing recurrent fallers were used as outcomes. Data of studies were analyzed separately and together.

Results: NEAT and ANFIS had better performance criteria compared to other models. The highest performance criteria were reported with NEAT when using PCR4 database and falls ≥1, and with both NEAT and ANFIS when pooling data together and using falls ≥2. However, sensitivity and specificity were unbalanced. Sensitivity was higher than specificity when identifying fallers, whereas the converse was found when predicting recurrent fallers.

Conclusions: Our results showed that NEAT and ANFIS were non-linear statistical models with the best performance criteria for the prediction of falls but their sensitivity and specificity were unbalanced, underscoring that models should be used respectively for the screening of fallers and the diagnosis of recurrent fallers.