This retrospective study used real-world inpatient administrative data from a tertiary hospital in Shanghai, China. The dataset contained 65,215 hospitalization records from January 1, 2023 to December 31, 2024. Each record represented a single inpatient episode and included demographic information, admission and discharge dates, primary diagnosis, principal procedure, DRG code, hospital charges, and length of stay.

This study was based on retrospective, fully de-identified administrative inpatient data from a tertiary hospital in Shanghai. No identifiable personal information was accessed. According to institutional policy, analyses of anonymized secondary administrative data are exempt from formal ethics committee review. Therefore, ethical approval and informed consent were not required.

A total of 65,215 inpatient episodes from January 2023 to December 2024 were initially identified.

To ensure complete follow-up for 30-day readmission, all discharges occurring in December 2024 were excluded (n = 3,109), yielding 62,106 hospitalizations for descriptive statistics and estimation of the overall readmission rate.

For multivariable modeling, a complete-case analysis was conducted, and 21 records with missing DRG-related variables were removed, resulting in 62,085 admissions included in the final logistic regression model. The proportion of missing data was 0.03% (21/62,106), which is negligible. Given this extremely low level of missingness, the impact of complete-case analysis on model estimates is unlikely to be substantial.

Thirty-day readmission was defined as any subsequent hospitalization recorded in the same hospital within 30 days after discharge from the index admission for the same patient identifier. Because the administrative database did not contain reliable markers distinguishing planned from unplanned admissions, all subsequent admissions within the 30-day window were included in the definition (12).

Hospitalization records were sorted by patient identifier and discharge date. For each admission, the next admission date (if any) was compared with the discharge date to determine whether a readmission occurred within 30 days. A binary variable was created (1 = readmission within 30 days, 0 = no readmission). Readmissions were captured only if they occurred within the same hospital, as cross-institution linkage was not available. Therefore, the estimated readmission rate should be interpreted as an intra-hospital readmission measure rather than a system-wide population rate.

The overall 30-day readmission rate in the complete follow-up cohort (n = 62,106) was 13.0% (8,104/62,106).

To assess potential overfitting and quantify optimism in model discrimination, we performed internal validation using bootstrap resampling (1,000 iterations). In each bootstrap sample, the model was refitted and evaluated, and optimism was estimated as the mean difference between bootstrap and original AUC values. An optimism-corrected AUC was then calculated. It is reported in the Results section.

All analyses were performed using SPSS 26.0 and R 4.3.3.

A two-sided p < 0.05 was considered statistically significant.

A total of 62,106 hospitalization records were included in the cohort used to estimate the 30-day readmission rate. Baseline characteristics stratified by readmission status are summarized in Table 1. Among all admissions, 54,002 (87.0%) were not followed by a 30-day readmission, whereas 8,104 (13.0%) resulted in an unplanned readmission within 30 days.

Patients who were readmitted tended to be older, with a median age of 64.1 years (IQR 52.8–70.8) compared with 62.7 years (IQR 47.3–70.8) among those not readmitted. In unadjusted comparisons, the readmission group had shorter lengths of stay (median 3 vs. 4 days) and lower hospitalization costs (median 9,634 vs. 15,208 CNY). They were also less likely to have undergone surgery (28.3% vs. 64.6%). The distribution of discharge year differed significantly between groups (p = 0.001).

A total of 8,104 unplanned readmissions occurred within 30 days, corresponding to an overall 30-day readmission rate of 13.0% in the cohort with complete follow-up (8,104/62,106). As shown in Table 1, patients who were readmitted tended to be older and had shorter lengths of stay, lower hospitalization costs, and a lower likelihood of undergoing surgery compared with those who were not readmitted.

Multivariable logistic regression results are presented in Table 2. After adjustment for age, length of stay, total cost, year of discharge, and DRG group, longer hospitalization duration was independently associated with a higher likelihood of 30-day readmission (OR 1.016 per day; 95% CI 1.009–1.023; p < 0.001). Total hospital cost showed a statistically significant but clinically negligible association with readmission (OR 1.000 per 1 CNY increase, p = 0.003). Although statistically detectable in this large sample, the effect size was extremely small, and even substantial increases in cost would correspond to minimal changes in predicted probability. This finding suggests that cost, when modeled at the individual level in its original scale, has limited clinical relevance as an independent predictor. When interpreted over a 1,000 CNY increase, the corresponding odds ratio remained very close to unity, indicating minimal clinical impact despite statistical significance.

Age and discharge year (2024 vs. 2023) were not statistically significant predictors. In contrast, DRG groups showed a strong overall association with readmission risk (global p < 0.001), indicating substantial heterogeneity across clinical case-mix categories. To illustrate the magnitude of between-DRG heterogeneity, we present the adjusted odds ratios for the five DRG categories with the highest and lowest readmission risks in Table 3. The highest-risk DRGs demonstrated more than five-fold variation in adjusted odds compared with the reference category, whereas the lowest-risk DRGs showed substantially reduced readmission probabilities. These findings underscore the structural variability in readmission risk across clinical pathways.

Although patients who were readmitted had shorter median lengths of stay in unadjusted comparisons (3 vs. 4 days; p < 0.001), the direction of association reversed in the fully adjusted model. After accounting for DRG category and other covariates, longer length of stay was associated with higher readmission risk (OR 1.016 per additional day; 95% CI 1.009–1.023; p < 0.001). This pattern suggests that, within the same DRG-defined clinical pathway, prolonged hospitalization likely reflects greater clinical instability, treatment complexity, or unresolved illness, thereby increasing the probability of early readmission.

To assess the linearity assumption for length of stay, a sensitivity analysis was performed by categorizing LOS into quartiles and re-estimating the model. The overall association between LOS categories and 30-day readmission remained statistically significant (global p = 0.004). Compared with the reference quartile, the highest LOS quartile demonstrated a statistically significant difference in readmission risk, while intermediate quartiles showed no consistent directional pattern. These findings support the robustness of the primary linear specification and suggest that modeling LOS as a continuous predictor did not materially distort the observed association.

The logistic regression model demonstrated acceptable discriminatory ability for predicting 30-day readmission. As shown in Figure 1, the area under the ROC curve (AUC) was 0.743 (95% CI 0.736–0.749), indicating moderate predictive accuracy.

The optimal probability threshold identified using Youden’s index was 0.167, corresponding to a sensitivity of 0.64 and a specificity of 0.79. These results suggest that the model is reasonably effective in distinguishing patients who will be readmitted from those who will not.

Bootstrap internal validation showed minimal optimism. The optimism-corrected AUC (0.743) was nearly identical to the apparent AUC, indicating minimal overfitting and stable internal performance.

The Hosmer–Lemeshow (HL) test was statistically significant (p < 0.001). However, given the large sample size (n = 62,085), even small deviations between observed and predicted probabilities may result in statistical significance. Therefore, HL test results should not be interpreted as evidence of poor calibration in isolation. Visual inspection of the decile-based calibration plot (Figure 2) demonstrated generally good agreement between predicted and observed readmission risks, with only mild deviation in the highest-risk groups. The Brier score was 0.098, indicating good overall accuracy of predicted probabilities.

The distribution of predicted probabilities stratified by readmission status is shown in Figure 3. Patients who were readmitted exhibited a visibly higher density peak at greater predicted probabilities, indicating that the model effectively assigns higher risk scores to individuals who ultimately experienced a 30-day readmission.

To assess clinical risk stratification capability, predicted probabilities were grouped into deciles. Observed 30-day readmission rates rose progressively across deciles (Figure 4), increasing from approximately 8% in the lowest decile to over 45% in the highest decile. This monotonic pattern demonstrates strong stratification performance and supports the potential utility of the model for identifying high-risk patients.

The overall 30-day readmission rate of 13.0% in this study is broadly consistent with international benchmarks for mixed-condition adult populations, although it remains lower than the condition-specific rates commonly reported in Medicare data for high-risk diagnoses such as heart failure or pneumonia. Age and length of stay emerged as the primary predictors of early readmission, aligning with patterns documented in multiple OECD health systems (13), where older patients—often with multimorbidity, polypharmacy, frailty, or functional decline—face a heightened likelihood of clinical deterioration shortly after discharge (14).

A key finding concerns the complex role of length of stay. In crude comparisons, patients who were readmitted had shorter hospitalizations, echoing concerns raised in the literature that prospective payment systems may be associated with organizational pressures toward earlier discharge timing, although our findings cannot establish causal relationships (15). However, after adjusting for DRG category and other covariates, the association reversed: within any given DRG-defined clinical pathway, longer stays were associated with a higher readmission risk. This apparent contradiction underscores that length of stay captures different dimensions of severity at different analytic levels. Across DRGs (“between-group”), shorter stays may reflect premature discharge. Within DRGs (“within-group”), longer stays likely signal unresolved clinical instability, treatment delays, postoperative complications, or complex therapeutic needs—factors that naturally increase the likelihood of returning within 30 days. Taken together, these findings should be interpreted as associations observed within a specific institutional context rather than as direct causal effects of DRG payment.

Hospitalization cost showed a statistically significant but clinically negligible association with readmission. Given the very small marginal effect observed, cost should not be interpreted as a meaningful independent driver of readmission risk, but rather as a potential proxy for unmeasured clinical complexity. Although the odds ratio per 1 CNY increase was statistically significant, the magnitude of effect was extremely small. Even when scaled to a 1,000 CNY increase, the change in predicted probability remained minimal, underscoring that statistical significance does not necessarily imply meaningful clinical or policy relevance. This is expected in a DRG-based system, where payments—not actual expenditures—drive hospital financial behavior. Cost therefore appears to offer limited clinical interpretability within this administrative framework, and our findings reinforce arguments that quality monitoring under DRG reform should shift away from cost-centered indicators toward patient-centered outcomes such as avoidable readmission, discharge safety, and care-transition quality.

The inclusion of DRG groupings was associated with improved model fit and discrimination. At the meantime, it revealed marked heterogeneity in readmission risk across clinical categories. These adjusted estimates reflect comparisons relative to the reference DRG group rather than absolute measures of intrinsic clinical severity. High-risk DRGs—including those involving acute medical instability or surgical episodes with intrinsic susceptibility to complications—displayed notably higher adjusted odds of readmission, whereas stable chronic or elective-procedural DRGs showed substantially lower risk. This pattern closely mirrors findings reported in the United States, England, Germany, and South Korea, indicating that variability in readmission risk is a structural feature of modern hospital systems rather than an artifact of local practice. The modest calibration challenges observed in our model (as reflected in the Hosmer–Lemeshow test) further emphasize the inherent difficulty of fitting a single universal prediction model across highly heterogeneous DRG populations.

Extensive international evidence has documented that DRG-based prospective payment systems have been associated with improvements in hospital efficiency and cost containment, may also generate unintended consequences for care continuity and post-discharge outcomes (15, 16). Experience from the United States, where the Hospital Readmission Reduction Program (HRRP) has linked financial penalties to early readmissions since 2012, illustrates both the potential and the limitations of readmission-based quality incentives (17, 18). HRRP has been associated with measurable reductions in readmission rates for targeted conditions (19), but subsequent evaluations raised concerns regarding risk-shifting behaviors, including increased use of emergency observation units, delayed readmissions beyond the 30-day window, and challenges in adequately adjusting for socioeconomic and clinical complexity (20). In England, the National Health Service initially piloted readmission penalties within the Payment by Results (PbR) framework but gradually moved toward blended payment models and integrated care pathways after evidence accumulated that punitive financial mechanisms alone were insufficient to address fragmentation between inpatient and community services (21). Similarly, health systems in Germany and Australia monitor readmissions closely but rely less on direct penalties, emphasizing instead the importance of standardized care pathways, strengthened transitional care, and collaboration between hospitals and primary care providers (22, 23). Across OECD countries, a consistent narrative has emerged: while DRG payment can improve efficiency, its influence on post-acute outcomes depends on the degree to which health systems support coordinated care and mitigate incentives for premature discharge.

China’s implementation of DRG-based payment differs from these international systems in both timing and policy orientation. Rather than introducing explicit readmission penalties, China’s reform has focused on standardizing clinical pathways, controlling inpatient expenditure growth, and establishing budget accountability across hospitals (24). In this context, our findings highlight how early readmission can serve as a sensitive indicator of the tensions inherent in a maturing DRG environment. The association between shorter length of stay and elevated unadjusted readmission risk suggests that hospitals may be operating under meaningful time and efficiency pressures. At the same time, the substantial variation in readmission across DRG categories underscores the need for condition-sensitive policy design, aligning closely with international experiences that emphasize differential management of high-risk diagnostic groups.

Overall, the patterns observed in this study reinforce global lessons: DRG-based payment reforms must be paired with robust transitional care, adequate risk adjustment, and system-level coordination to avoid unintended quality trade-offs. China’s ongoing reform presents a critical opportunity to incorporate these lessons early in the development of its national DRG framework.

The implications of our findings extend beyond simple risk prediction and touch on the broader architecture of China’s ongoing DRG-based payment reforms. First, our results highlight the value of incorporating readmission indicators into routine quality monitoring within DRG systems. Because DRGs primarily incentivize cost containment and efficiency, hospitals may unintentionally shorten inpatient stays in ways that are associated with higher observed readmission rates. A risk-adjusted readmission measure—calculated either across all DRGs or within specific high-risk groups—could serve as a counterbalancing metric. Integrating such indicators into performance dashboards or payment adjustments would help ensure that efficiency gains do not come at the expense of clinical safety. However, these policy considerations should be interpreted in light of the study’s single-center design, and broader national implementation would require validation across diverse hospital settings and regional DRG contexts.

Second, the pronounced variation in readmission risk across major DRG categories suggests that China would benefit from differentiated DRG-level policy levers, rather than uniform rules. In mature DRG systems such as Germany and Australia, high-severity or clinically volatile DRGs are often granted extended length-of-stay thresholds, enhanced payment flexibility, or supplemental quality monitoring (25, 26). Adopting similar nuance in Chinese DRGs—such as tailoring outlier policies or providing higher case weights for clinically unstable DRGs—may help reduce pressure toward premature discharge and better align reimbursement with patient needs.

Third, our findings underscore the need to strengthen post-acute continuity of care, given that a considerable proportion of readmissions stem from failures in transitional care, medication management, and outpatient follow-up. International evidence has consistently shown that structured transition programs, telehealth-assisted monitoring, primary care linkage, and community-based rehabilitation can effectively lower readmission rates (27, 28). In China, where digital infrastructure is rapidly advancing and community health centers are expanding under hierarchical medical reform, integrating such transitional care models into DRG payment rules may yield substantial quality and efficiency benefits.

Fourth, the calibration limitations of our prediction model reveal the constraints of relying solely on administrative data for risk adjustment. Important predictors—frailty, functional status, socioeconomic vulnerability, caregiver availability, and health literacy—remain unmeasured in most Chinese hospital information systems. As China develops a unified national health information platform, incorporating structured clinical data, medication histories, and patient-reported outcomes will be crucial for improving risk adjustment and building fairer benchmarking frameworks under DRG payment.

Finally, our findings reinforce that readmission risk is not merely a reflection of inpatient decision-making but rather a system-level indicator shaped by broader care fragmentation, uneven access to primary care, and limited long-term care resources. Policymakers should therefore interpret readmission metrics as part of an integrated ecosystem of incentives and constraints, recognizing that hospitals alone cannot rectify upstream social and outpatient challenges.

Operationally, the proposed model could be integrated into hospital information systems as part of routine quality monitoring dashboards. For example, predicted readmission risk scores could be generated at discharge to flag high-risk patients for enhanced transitional care planning, medication reconciliation, or early follow-up scheduling. At the system level, payers could use risk-adjusted readmission benchmarks within specific high-risk DRG categories to identify outlier institutions and guide targeted quality-improvement interventions rather than applying uniform penalties. Such applications should prioritize risk adjustment, transparency, and supportive feedback mechanisms, ensuring that predictive analytics are used to improve care coordination rather than to impose simplistic performance sanctions.

This study has several notable strengths. It draws on a large, real-world administrative dataset comprising more than 62,000 inpatient encounters from a high-volume tertiary hospital, enabling a granular assessment of readmission patterns across diverse DRG categories. The analytic strategy incorporated multiple complementary performance metrics—including discrimination, calibration, and risk stratification—providing a comprehensive evaluation rarely seen in existing Chinese DRG literature. Furthermore, by explicitly comparing crude and adjusted findings for length of stay, the study offers nuanced insights into how DRG payment structures shape clinical decision-making at both the between-DRG and within-DRG levels.

Nonetheless, several limitations must be acknowledged. First, the analysis relied exclusively on hospital administrative data, which lack important clinical detail such as comorbidity scores, physiological indicators, disease severity, frailty status, and functional capacity. These omissions may have led to residual confounding and restricted the ability of the model to fully capture patient complexity. Second, because readmissions were identified only within the same hospital, cross-institution readmissions were missed. International evidence suggests that 10–30% of all unplanned readmissions may occur at different facilities (29), which implies that our readmission rate is likely an underestimate. Third, the absence of outpatient, pharmacy, and primary-care data limits our ability to account for transitional care processes and post-discharge adherence, both of which are major contributors to early readmissions. Fourth, as a single-center study conducted within a specific institutional and regional context, generalizability is limited; patterns observed in Shanghai may differ substantially from those in less-resourced provinces or hospitals with different DRG maturity levels. Moreover, because planned and unplanned readmissions could not be distinguished in the data, the reported readmission rate may be upwardly biased, and some degree of nondifferential misclassification is possible. Finally, although complete-case analysis was applied, the proportion of missing data was extremely small (0.03%), reducing the likelihood of meaningful bias. Nevertheless, if missingness was related to documentation practices or unobserved severity, minor selection effects cannot be entirely excluded.

Future research should build upon these findings by integrating more granular clinical and patient-reported data to improve risk adjustment and enhance the interpretability of readmission models in DRG environments. Linking inpatient data with primary care, pharmacy claims, and community health records would allow for a more complete characterization of transitional care gaps and post-discharge trajectories. Prospective cohort studies, particularly those incorporating frailty measures, functional assessments, biomarkers, and socioeconomic indicators, could provide deeper insight into mechanisms underlying early readmission. Additionally, multi-center investigations across provinces with different DRG implementation stages would help assess external generalizability and identify region-specific policy levers. As China advances toward a national health information infrastructure, future studies should explore machine-learning approaches and dynamic risk monitoring that support real-time, individualized prediction. Evaluating the impact of targeted interventions—such as structured discharge planning, telehealth follow-up, and community-based care coordination—will also be critical for translating predictive insights into measurable improvements in patient outcomes.