Development and validation of dual-time-window readmission prediction models under China’s medical insurance multi-payment reform

Background Readmission is an important indicator of healthcare quality, reflecting continuity of care, treatment outcomes, and the quality of discharge management. However, existing readmission prediction studies have primarily focused on 30-day readmission as the primary outcome, with limited attention to differences in the underlying risk mechanisms, predictive performance, and clinical applicability across different readmission time windows. Meanwhile, under the context of China’s multi-payment reform and the Coordinated Tripartite Medical Reform, limited evidence is available on the development of readmission prediction models based on real-world data and their application to healthcare quality monitoring and the management of high-risk patients. Therefore, this study aimed to develop and validate 15-day and 30-day all-cause same-hospital readmission prediction models using real-world hospitalization data, compare the predictive factors and model performance across different readmission time windows, and provide empirical evidence to support healthcare quality monitoring and risk management under the ongoing payment reform. Methods A single-center retrospective design was adopted, and data from 57,677 hospitalizations at a large tertiary hospital in China between January 2021 and July 2025 were collected. Multivariable logistic regression models were developed for 15-day and 30-day all-cause same-hospital readmissions, respectively. Model performance was evaluated using the AUC, Brier score, calibration plots, bootstrap internal validation, and temporal external validation. Differences in predictor associations between the two readmission time horizons were further examined. Results The 15-day and 30-day readmission rates were 9.47 and 10.92%, respectively. The 15-day model achieved an AUC of 0.840 (95% CI: 0.828–0.843) and a Brier score of 0.0620; the 30-day model achieved an AUC of 0.820 (95% CI: 0.810–0.825) and a Brier score of 0.0735. Both AUCs were statistically significant. Both models demonstrated good discrimination and calibration, with the 15-day model consistently outperforming the 30-day model across all performance metrics. In the temporal external validation, the model achieved an AUC of 0.810 and a Brier score of 0.136, indicating good generalizability. Age, LOS, surgical status, and number of secondary diagnoses were independently associated with higher odds of readmission (overall p < 0.05), while higher total cost was associated with lower odds of readmission. Risk stratification showed that the observed readmission rate exceeded 58% in the highest-risk decile and was below 2% in the lowest-risk group. Conclusion This study developed and validated dual-time-horizon readmission prediction models based on real-world data, extending the conventional approach of single-horizon readmission risk assessment. The models provide a methodological framework for risk stratification and healthcare quality monitoring across different post-discharge periods. They may support early discharge risk screening, facilitate prioritization of transitional care resources for high-risk patients, and provide empirical evidence to inform quality assessment, risk adjustment, and management decision-making under the multi-payment reform and the Coordinated Tripartite Medical Reform.