Identifying risk prediction models and predictors for hospital readmission in patients with medical conditions: A systematic review and meta-analysis

BACKGROUND: Hospital readmission is a frequent adverse outcome among patients with medical conditions, with approximately 20% being readmitted within 30 days of discharge. However, the factors that significantly influence early readmission remain poorly understood, and it is unclear how to comprehensively evaluate predictors to identify patients at higher risk of readmission effectively.

OBJECTIVE: To identify internally validated prediction models for all-cause readmission within 28-31 days of patients with medical conditions and to summarize the types of candidate and final predictors as well as the predictive performance of the models.

DESIGN: Systematic review and meta-analysis of observational studies.

METHODS: Electronic databases (CINAHL, MEDLINE, and EMBASE) were searched until January 2024, along with the reference lists on the included studies from the search. Data from the included studies were extracted using the CHARMS checklist for prediction models. The PROBAST toll for prediction models assessed the risk of bias and applicability.

RESULTS: 24,322 studies were retrieved, and after the selection process, 16 prediction studies were included in the review. 12 of the studies were retrospective cohorts and exclusively used administrative data. The most commonly reported predictors with a significant impact on 28-31 day readmissions are age, higher Charlson index score, congestive heart failure, chronic obstructive lung disease, chronic renal insufficiency, arrhythmia and atrial fibrillation, length of stay, emergency department visits within six months, number of admissions last year, cancer and oncology services, polypharmacy, low sodium level, low hemoglobin level, and lower albumin level. Most studies had a high risk of bias, primarily in the analysis domain. 13 models reported the AUC, and the pooled AUC value was 0.71 (0.68, 0.74), indicating a moderate performance.

CONCLUSION: Although most of the included studies demonstrated moderate to good discrimination, many models exhibited a high overall risk of bias. Assessing key predictors can be challenging, as they are often not routinely captured in administrative data.

REGISTRATION: The protocol was registered in Open Science Framework (OSF) on March 2024 (doi.org/10.17605/OSF.IO/PDSH5).