New predictive models for falls among inpatients using public ADL scale in Japan: A retrospective observational study of 7,858 patients in acute care setting

Aim: Most predictive models for falls developed previously were awkward to use because of their complexity. We developed and validated a new easier-to-use predictive model for falls of adult inpatients using easily accessible information including the public ADL scale in Japan. Methods: We retrospectively analyzed data from Japanese adult inpatients in an acute care hospital from 2012 to 2015. Two-thirds of cases were randomly extracted to the test set and one-third to the validation set. Data including age, sex, activity of daily living (ADL), public scales in Japan of ADL “bedriddenness rank,” and cognitive function in daily living, hypnotic medications, previous falls, and emergency admission were derived from hospital records. Falls during hospitalization were identified from incident reports. Two predictive models were created by multivariate analysis, each of which was assessed by area under the curve (AUC) from the validation set. Results: A total of 7,858 adult participants were available. The AUC of model 1, using 13 factors—age, sex (male), emergency admission, use of ambulance, referral letter, admission to Neurosurgery, admission to Internal Medicine, use of hypnotic medication, permanent damage by stroke, history of falls, visual impairment, independence of eating, and bedriddenness rank—with low mutual collinearity and showing significant relationship by multivariate logistic regression analysis, was 0.789 in the validation set. The AUC of parsimonious model 2, using age and seven factors—sex (male), emergency admission, admission to Neurosurgery, use of hypnotic medication, history of falls, independence of eating, and bedriddenness rank—showing statistical significance by multivariate analysis in model 1, was 0.787 in the validation set. Conclusions: We proposed new predictive models for inpatients’ fall using the public ADL scales in Japan, which had a higher degree of usability because of their use of simpler and fewer (8 or 13) predictors, especially parsimonious model 2.