Abstract WP134: Predictive Models for 30-Day Readmission After Stroke: A Systematic Review and Meta-Analysis

Background: Stroke readmissions highlight gaps in post-discharge care that require urgent action. In the United States, 18.2% of stroke patients are readmitted within 30 days of discharge, with a 12.4% readmission rate for ischemic stroke. While several predictive models have been developed to identify factors associated with readmission, evidence on their performance and clinical utility remains limited We aimed to evaluate and compare the performance of logistic regression and machine learning models for predicting all cause and stroke-specific 30-day readmissions using data from existing literature. Methods: Following the PRISMA guidelines, we systematically searched PubMed, Embase, Web of Science, and Google Scholar for studies published between January 2021 and July 2025. Of 292 records identified, 13 studies met inclusion criteria. We evaluated model quality using rehospitalization rates and AUC, performed subgroup analysis comparing of all-cause (AC) versus stroke-specific (SS) readmissions using logistic regression (4) and machine learning methods (13) approaches. Results: Total number of subjects were 203,399 (SS readmissions = 280; AC readmissions = 20340). 17 models provided AUCs for meta-analysis: SS readmission models (n = 2; pooled readmission rate: 2.0%), and AC readmission models (n = 15; readmission rate: 8.0%). Chen et.al.2022 demonstrated the highest AUCs for AC readmissions (0.94 and 0.89) despite low readmission rates, followed by Hu et al. (2025) with an AUC of 0.91 and Lv et al. (2023) with an AUC of 0.80. The overall readmission proportion across all studies was 11% (95% CI: 7%–16%), with a weighted combined proportion of 6.8%, based on subgroup contribution of 80% AC and 20% SS. The AUC across studies was 0.72 (95% CI: 0.67–0.78). Subgroup-specific AUCs were 0.67 (95% CI: 0.54–0.79) for SS and 0.70 (95% CI: 0.69–0.70) for AC. (Figure) Between-study heterogeneity was considerable (Q = 829.35, p < 0.001; I 2 =98.07%; τ 2 = 0.01). Conclusion: Our results suggest that overall predictive accuracy of 30-day readmissions remain limited, regardless of strong potential of some models. The high level of heterogeneity indicates that methodological and population differences influence performance, emphasizing the need for external validation and standardization of modeling approaches. Future studies should focus on developing robust, generalizable readmission risk models refined and multi-site validation and transparent reporting standards.