Thirty-Day Readmission Risk Stratification at Hospital Discharge: Single-Center Temporal Validation and Nursing Implementation Readiness Study

Abstract Background Discharge-planning nurses and case managers need to prioritize transitional care, but predictive ranking alone does not identify which patients will benefit from a specific intervention. Objective This study aimed to develop and temporally validate a local discharge time model for 30-day readmission; evaluate calibration, risk tiers, clinical utility, and subgroup performance; and describe historical nursing action cohorts only as non–model-triggered feasibility evidence for a future pathway. Methods This single-center observational study used 41,935 discharge episodes from 2017 to 2018. After excluding 2.3% (964/41,935) in-hospital deaths, 48.5% (19,890/40,971) of episodes from 2017 were used for development, and 51.5% (21,081/40,971) of episodes from 2018 were used for temporal validation. The outcome was the same-institution all-cause readmission within 30 days, defined as the union of readmissions on days 1 to 14 and 15 to 30. Logistic regression, random forest, and Extreme Gradient Boosting were calibrated using 5-fold out-of-fold predictions from 2017. The 95% CIs were obtained with 1000 stratified bootstrap replicates. The separate historical cohorts included shared decision-making (n=72), long-term care linkage (n=364), and remote monitoring (n=43). None was selected or routed by model output. Results The validation cohort had 20.7% (4361/21,081) readmissions. Random forest was selected from development cross-validation and achieved an area under the receiver operating characteristic curve (AUROC) of 0.816 (95% CI 0.809-0.823), area under the precision-recall curve (AUPRC) of 0.593 (95% CI 0.579-0.607), and Brier score of 0.119 (95% CI 0.116-0.121); the calibration intercept was –0.117, the slope was 0.974, and the observed-to-expected ratio was 0.951. Readmission occurred in 6.5% (661/10,120) of low-risk episodes, 18.5% (1160/6271) of moderate-risk episodes, 47.3% (1769/3739) of high-risk episodes, and 81.1% (771/951) of very high–risk episodes. The principal diagnosis group accounted for 84.6% of the random forest importance. After excluding index episodes labeled as planned admission or discharge and refitting, the AUROC decreased to 0.726, and the AUPRC decreased to 0.284. In historical cohorts, discharge direction changed in 41.7% (30/72) of cases; 82.7% (301/364) of patients used long-term care; 90.2% (268/297) of patients with valid dates started service within 7 days; and the uncontrolled remote monitoring cohort showed exploratory 3-month reductions in hospital days, emergency visits, and readmissions. Conclusions The recalculated local model showed useful temporal discrimination and near-unity calibration after probability calibration, but its performance depended heavily on diagnosis coding and potentially planned readmissions. External generalizability and model actionability remain unestablished. Prospective silent-run evaluation, local threshold review, subgroup monitoring, and controlled testing of the nurse-led response are required before clinical activation.