Patient Ambulations Predict Hospital Readmission

Abstract

Improved functional ability and physical activity are strongly associated with a broad range of positive health outcomes including reduced risk of hospital readmission. This study presents an algorithm for detecting ambulations from time-resolved step counts gathered from remote monitoring of patients receiving hospital care in their homes. It examines the statistical power of these ambulations in predicting hospital readmission. A diverse demographic cohort of 233 patients of age 70.5±16.8 years are evaluated in a retrospective analysis. Eleven statistical features are derived from raw time series data, and their F-statistics are assessed in discriminating between patients who were and were not readmitted within 30 days of discharge. Using these features, logistic regression models are trained to predict readmission. The results show that the fraction of days with at least one ambulation was the strongest feature, with an F-statistic of 17.2. The models demonstrate AUROC performances of 0.741, 0.766 and 0.769 using stratified 5-fold train-test splits in all included patients (n=233), congestive heart failure (CHF, n=105) and non-CHF (n=128) patient subgroups, respectively. This study suggests that patient ambulation metrics derived from wearable sensors can offer powerful predictors of adverse clinical outcomes such as hospital readmission, even in the absence of other features such as physiological vital signs.Index Terms-readmission, ambulation, step count, heart failure, physical activity, regression, actigraphy, accelerometer.