Predicting falls within 3 months of emergency department discharge among community-dwelling older adults using self-report tools versus a brief functional assessment.

ABSTRACT

BACKGROUND: Identifying older adults with risk for falls prior to discharge home from the Emergency Department (ED) could help direct fall prevention interventions, yet ED-based tools to assist risk stratification are under-developed. The aim of this study was to assess the performance of self-report and functional assessments to predict falls in the 3 months post-ED discharge for older adults. METHODS: A prospective cohort of community-dwelling adults age 60 years and older were recruited from one urban ED (N = 134). Participants completed a single item screen for mobility (SIS-M), the 12-item Stay Independent Questionnaire (SIQ-12), and the Timed Up and Go test (TUG). Falls were defined through self-report of any fall at 1- and 3-months and medical record review for fall-related injury 3-months post-discharge. We developed a hybrid-convolutional recurrent neural network (HCRNN) model of gait and balance characteristics using truncal 3-axis accelerometry collected during the TUG. Internal validation was conducted using bootstrap resampling with 1000 iterations for SIS-M, FRQ, and GUG and leave-one-out for the HCRNN. We compared performance of M-SIS, FRQ, TUG time, and HCRNN by calculating the area under the receiver operating characteristic area under the curves (AUCs). RESULTS: 14 (10.4%) of participants met our primary outcome of a fall or fall-related injury within 3-months. The SIS-M had an AUC of 0.42 [95% confidence interval (CI) 0.19-0.65]. The SIQ-12 score had an AUC of 0.64 [95% confidence interval (CI) 0.49-0.80]. The TUG had an AUC of 0.48 (95% CI 0.29-0.68). The HCRNN model using generated accelerometer features collected during the TUG had an AUC of 0.99 (95% CI 0.98-1.00). CONCLUSION: We found that self-report and functional assessments lack sufficient accuracy to be used in isolation in the ED. A neural network model using accelerometer features could be a promising modality but research is needed to externally validate these findings.