Validation of two risk-prediction models for recurrent falls in the first year after stroke: a prospective cohort study

Background: Several multivariable models have been derived to predict post-stroke falls. These require validation before integration into clinical practice. The aim of this study was to externally validate two prediction models for recurrent falls in the first year post-stroke using an Irish prospective cohort study. Methodology: Stroke patients with planned home-discharges from five hospitals were recruited. Falls were recorded with monthly diaries and interviews six and 12 months post-discharge. Predictors for falls included in two risk-prediction models were assessed at discharge. Participants were classified into risk-groups using these models. Model 1, incorporating inpatient falls-history and balance, had a six-month outcome. Model 2, incorporating inpatient near-falls history and upper limb function, had a twelve-month outcome. Measures of calibration, discrimination (area under the curve (AUC)) and clinical utility (sensitivity/ specificity) were calculated. Results: 128 participants (mean age=68.6 years, SD=13.3) were recruited. The fall status of 117 and 110 participants was available at six and 12 months respectively. Seventeen and 28 participants experienced recurrent falls by these respective timepoints. Model 1 achieved an AUC=0.56 (95% CI 0.46–0.67), sensitivity=18.8% and specificity=93.6%. Model 2 achieved AUC=0.55 (95% CI 0.44–0.66), sensitivity=51.9% and specificity=58.7%. Model 1 showed no significant difference between predicted and observed events (Risk Ratio (RR)=0.87, 95% CI 0.16–4.62). In contrast, model 2 significantly over-predicted fall events in the validation cohort (RR=1.61, 95% CI 1.04–2.48). Conclusions: Both models showed poor discrimination for predicting recurrent falls. A further large prospective cohort study would be required to derive a clinicallyuseful falls-risk prediction model for a similar population.