Predicting unplanned acute care transfer from an inpatient rehabilitation facility across rehabilitation populations

Abstract

Risk factors for acute care transfers (ACTs) from inpatient rehabilitation facilities (IRFs) vary by setting and patient population. Prior risk prediction models designed to apply to all IRF patients in an IRF have achieved poor to acceptable predictive value, while better performance (C-statistic > 0.8) was attained by models isolated to specific impairment groups. Our aim was to determine if a general risk prediction model for ACT attained sufficient predictive utility when applied across all patients requiring IRF stay. We retrospectively reviewed adults admitted to an academic, rural IRF between January 2016 and September 2019. Predictors were selected based on variables that were statistically significant in prior studies, and covariates significant on bivariate analysis were included in a multivariable logistic regression model of ACT. The predictive utility of the model (bootstrapped C-statistic) was evaluated within each impairment group with greater than 100 patients. The analysis included 2275 patients (median age: 64 years; 57% male; 7% with unplanned ACT). Multivariable logistic regression of unplanned ACT demonstrated fair predictive utility [C-statistic: 0.75; 95% confidence interval (CI): 0.72–0.79], while applying the same risk score to individual impairment groups revealed highly heterogeneous C-statistics, from 0.81 (95% CI: 0.72–0.91) among patients with spinal cord dysfunction to 0.63 (95% CI: 0.50–0.75) among patients with amputation. Our derivation of a comprehensive risk score for ACT based on all patients admitted to IRF demonstrated fair predictive utility in the overall cohort, but predictive utility differed strongly across impairment groups. Further work is needed to develop personalized prediction of ACT risk based on the intersection of impairment group, clinical status at IRF admission, and pertinent medical and surgical history.