Regression-Based Balance Discordance Refines Prediction of Prospective Falls After Subacute Stroke: A Secondary Analysis of the LEAPS Trial

Objective: To determine whether continuous, regression-based measures of balance discordance refine prediction of prospective, recurrent falls after subacute stroke relative to a previously published threshold-based classification in the same cohort, and to characterize how a purely data-driven, dimensionality-reduction-based composite expresses the same underlying measures. Design: Secondary, observational analysis of a randomized controlled trial. Recurrent falls were modeled using Andersen-Gill Cox proportional hazards regression with robust standard errors clustered by participant. Setting: Five-site, National Institutes of Health/National Institute of Neurological Disorders and Stroke funded multicenter trial (Locomotor Experience Applied Post-Stroke, LEAPS). Participants: 384 adults with subacute stroke (of 408 randomized in the parent trial) with valid baseline balance confidence, walking, and balance scores and at least one recorded follow-up interval (2,357 person-intervals, 179 fall events). Main Outcome Measures: Prospective, recurrent falls over 12 months, ascertained via self-report falls diary. Balance discordance was computed as the residual from regressing balance confidence on objective walking distance, Berg Balance Scale performance, or both jointly (multivariable discordance); threshold-based classification used established cut-points on the same measures and principal component analysis served as secondary comparators. Results: Single-variable discordance was associated with lower fall hazard (hazard ratio [HR] = 0.79, 95% CI 0.63-0.99, p = .037), and multivariable discordance was associated with higher fall hazard (HR = 1.27, 95% CI 1.03-1.58, p = .026). None of three threshold-based classification indicators reached significance. Multivariable discordance and a three-component principal-component composite showed comparable model fit (concordance 0.616 vs. 0.618) and were substantially correlated (Rsquared = .49). Conclusions: Continuous, regression-based balance discordance predicted prospective, recurrent falls where a previously published threshold-based classification did not. The divergent direction of single- and multivariable findings suggests overconfidence relative to combined objective function, rather than confidence broadly, may be the more specific marker of fall risk.