Prediction of 180-Day Emergency Department Readmission: A Retrospective Study with Temporal Validation

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Background: Demographic change, worsening of general health risk profiles and lack of resources cause growing challenges in healthcare, especially in acute and emergency care. To explore the potential of digitally facilitated personalized patient flow management tools in acute care at Austrian emergency departments, we developed and validated a model to predict the individual probability of readmission after an initial acute treatment. Methods: For this exploratory study, a retrospective data analysis on 50,849 cases of acute treatment was conducted, which were retrievable from the routine dataset of the emergency department of the Medical University of Vienna at the tertiary care central hospital. The prediction model was developed based on cases documented between January 2013 and August 2018 and temporally validated using cases registered between August 2018 and April 2022. The prediction model was developed using multivariable Firth’s logistic regression and confirmed by temporal validation. Quality and usefulness were evaluated using calibration plots, Kaplan–Meier curves, C-statistics and decision curve analyses. Results: A total of 25,423 cases were analyzed in the development dataset (median age 59 years, IQR 41–73; 48.15% female), and 25,426 cases in the validation dataset (median age 51 years, IQR 32–70; 50.88% female). The incidence rates of readmission were 3.34 per 100 patient-years in the development dataset and 7.49 per 100 patient-years in the validation dataset. The final score included seven predictors: age > 50 years (“Yes”, 6 points), sex (“male”, 2 points), country of birth (“Austria”, 1 point), married/partnered (“Yes”, 2 points), place of residence Vienna (“Yes”, 5 points), internal medical reason for admission (“Yes”, 1 point), and clear diagnosis possible (“Yes”, 2 points). The final score showed good calibration (R2 = 0.952 and R2 = 0.891) and modest discrimination (C-indices: 0.60 and 0.62) in both the development and validation datasets. The decision curve analyses suggested clinical net benefit. The results of the sensitivity analyses indicated a robust model. Conclusions: A substantial rise in incidence rates indicates an urgent need for action. The present model showed good calibration but modest discrimination in predicting readmission after initial acute treatment at an emergency department. These findings support further refinement and external validation of the model before its potential use in digitally facilitated personalized patient flow management.