We performed a comprehensive search across multiple databases, including PubMed, Web of Science, The Cochrane Library, Embase, Cumulative Index to Nursing and Allied Health Literature (CINAHL), China National Knowledge Infrastructure (CNKI), Wanfang Database, China Science and Technology Journal Database (VIP), and SinoMed, covering all records from their establishment up to November 1, 2024. The search strategy targeted three primary concepts: PCI, readmission, and prediction. The core Boolean logic used was: “(PCI OR percutaneous coronary intervention) AND (readmission OR rehospitalization) AND (prediction OR prognostic OR risk model)”. Detailed, database-specific search strategies are provided in Supplementary Table 2.

The inclusion criteria were: (1) research involving individuals who underwent PCI; (2) observational cohort or case-control studies; (3) readmission post-PCI as the reported outcome; and (4) inclusion of a predictive model. Exclusion criteria encompassed: (1) research focusing solely on risk factors for readmission without developing risk prediction models; (2) studies lacking accessible full texts; (3) non-peer-reviewed materials, such as conference abstracts and agency reports; (4) redundant or overlapping publications; and (5) studies published in languages other than English or Chinese.

Two reviewers (MY and HW) independently screened articles according to the predefined inclusion criteria, with discrepancies resolved by a third reviewer (FH).

Data extraction was conducted using the CHARMS framework (see Supplementary Table 3). The extracted data from the selected studies were classified into four categories: (1) General study information, including the first author, year of publication, study design, data source, study period, and outcome definition; (2) Basic model information, encompassing sample size, outcome event rate, events per variable (EPV), model development method, variable selection method, handling of missing data, and processing of continuous variables; (3) Model performance, covering discrimination, calibration, type of validation, and formats for presenting the risk prediction models; and (4) Predictors, detailing the number of candidate variables and final predictors.

The risk of bias and applicability of the included studies were assessed using the PROBAST. The risk of bias assessment encompassed all four PROBAST domains: participant selection, predictors, outcome, and analysis. Results were stratified by domain, and the risk of bias was visualized using the robvis tool.

A meta-analysis was conducted to assess the discriminative performance of readmission prediction model, as measured by the area under the receiver operating characteristic curve (AUC). The analysis followed a structured process. First, to stabilize variance, AUC values were transformed to the logit scale using the formula: $logit(AUC)=\ln \left(\frac{AUC}{1-AUC}\right)$. The standard error was calculated using the expression: $S{E}_{logit(AUC)}=\sqrt{\frac{AUC(1-AUC)}{n\cdot {\varphi }^2}}$, where $\varphi$ denotes the probability density function at the optimal cutoff point. Effect sizes were pooled using the inverse-variance weighting method. Heterogeneity was evaluated using the Q-test and the I² statistic. Substantial heterogeneity was defined as I² $>$ 50% and Q-test p $\le$ 0.1, warranting the use of a random-effects model. In contrast, low heterogeneity (I² $\le$ 50% and Q-test p $>$ 0.1) supported the use of a fixed-effects model. For analyses with significant heterogeneity (I² $>$ 50%), subgroup analyses were conducted to explore potential sources of variability, stratified by publication date, study region, type of readmission, and modeling methods. Funnel plot asymmetry was assessed both visually and statistically using Egger’s linear regression test, with p 0.05 considered indicative of potential publication bias. All analyses were performed in R (version 4.4.1; R Foundation for Statistical Computing, Vienna, Austria) using the meta and metafor packages.

From the database searches, 1043 records were retrieved, of which 176 duplicates were excluded. Following the screening of 867 articles, 839 were excluded as irrelevant. A further 28 articles were eliminated for specific reasons: conference abstracts (n = 2), absence of a risk prediction model (n = 12), fewer than two predictors (n = 1), abstract only (n = 2), and non-primary literature (n = 1). In the end, 10 studies met the inclusion criteria, presenting a combined total of 18 prediction models for hospital readmission following PCI (Fig. 1).

The fundamental characteristics of the included studies are summarized in Table 1 (Ref. [17, 32, 33, 34, 35, 36, 37, 38, 39, 40]). These studies were published from 2013 and 2024, with six conducted in China and four in the United States. Among these, eight studies adopted a retrospective cohort design, while two followed a prospective cohort design. Seven studies recruited patients from a single center, while the remaining three reported results from two or more centers. Regarding study populations, nine investigations examined individuals with coronary heart disease, including two that specifically concentrated on elderly individuals aged 60 and older. One study focused on participants with acute coronary syndrome and type 2 diabetes mellitus. Sample sizes varied from 247 to 388,078 participants, with the mean or median age of participants spanning 59 to 73 years. The outcomes of interest included heart-related readmissions (n = 6), with subcategories such as myocardial infarction-related readmissions (n = 2), major adverse cardiac events (MACE)-related readmissions (n = 2), and congestive heart failure (CHF)-related readmissions (n = 1). Four studies investigated all-cause readmissions. Follow-up periods for readmission varied, with five studies evaluating 30-day readmissions and another five assessing 1-year readmissions. Reported readmission rates varied widely: 30-day all-cause readmission rates spanned from 7.94% to 11.36%, while 1-year heart-related readmission rates ranged from 16.93% to 31.44%.

The model information is detailed in Table 2 (Ref. [17, 32, 33, 34, 35, 36, 37, 38, 39, 40]) (The complete version is provided in Supplementary Table 4). Logistic regression was the primary method for model development, utilized in all included studies. Additionally, machine learning methods were applied in 2 of the 10 included studies. Specifically, 2 studies employed random forest (RF), while 1 study each used decision tree (DT), support vector machine (SVM), eXtreme Gradient Boosting (XGBoost), and Adaptive Boosting (AdaBoost). Fig. 2 summarizes all predictors included in the final models. Age was the most frequently used predictor, appearing in seven models. Other common predictors included heart failure, diabetes, chronic lung disease, and triple vessel lesion, each used in four models.

Model discrimination was reported in nine studies, with C-statistic values varying between 0.660 and 0.899. Calibration was assessed in seven studies, most commonly using the Hosmer-Lemeshow test.

Among the included studies, eight conducted internal validation, with three utilizing bootstrapping, two employing random splitting, one applying cross-validation, and one adopting temporal validation. One study did not specify the internal validation method used. Additionally, one study performed both internal and external validation. The complete set of model features is presented in Supplementary Table 4.

Table 3 (Ref. [17, 32, 33, 34, 35, 36, 37, 38, 39, 40]) and Fig. 3 summarize the risk of bias and applicability assessments for the included studies (The bias risk assessment criteria and procedures are detailed in Supplementary Table 5). Eight studies were deemed to have a high risk of bias, while four were identified to carry a high risk of applicability.

In the “participants” domain, all ten studies were classified as having a low risk of bias, as they either employed prospective cohort designs or retrospective cohorts that included all or consecutive PCI patients over a defined period. Within the “predictor” domain, all ten studies exhibited a low risk of bias. These were predominantly single-center studies, where commonly used predictors had standardized measurement protocols, and fixed variables like age and sex were inherently less susceptible to measurement bias. In the “outcome” domain, seven studies were also classified to have an unclear risk of bias because they failed to report whether outcomes and predictors were assesses independently (i.e., blind assessment) [32, 33].

In the “analysis” domain, eight model development studies were judged to have a high risk of bias. Specific issues included the following:

$\bullet$ Six studies had inadequate sample sizes, failing to meet the criterion of more than 10 EPV [17, 32, 33, 34, 35, 36].

$\bullet$ One study categorized continuous variables, potentially resulting in a loss of information [37].

$\bullet$ Three studies relied on univariate analysis for variable selection [33, 34, 36].

$\bullet$ Four studies did not comprehensively evaluated the predictive performance of their models [33, 35, 37, 38].

$\bullet$ Three study failed to address model overfitting, underfitting, or optimism in performance assessment [36, 37, 39].

$\bullet$ Five studies did not report the coefficients of predictors in their multivariate regression model [32, 35, 37, 39, 40].

$\bullet$ Two studies relied exclusively on internal validation using a single random split sample [37, 39].

$\bullet$ None of the studies reported complexities inherent in the data.

Two studies conducted external validation of previously developed predictive models. One study included at least 100 participants with outcome events in the validation cohort, meeting recommended sample size criteria. The other study did not report the number of outcome events but involved a large sample size (n = 2978), likely providing sufficient statistical power. Both studies applied the same dichotomization strategies and cut-off thresholds as defined in the original models.

For applicability, in the “patients” domain, three studies were classified as having a high risk of applicability because they focused on elderly patients with coronary heart disease and concurrent diabetes [33, 37, 38]. In the “outcome” domain, one study was assessed to have a high risk of applicability because the outcomes included not only readmission but also cardiovascular and all-cause mortality [35].

This study conducted a meta-analysis of relevant factors quantitatively synthesized from 10 studies, encompassing 14 predictors. Table 4 presents the primary meta-analysis results for all investigated predictors, including pooled effect sizes (odds ratios [ORs]) and heterogeneity statistics. Random- or fixed-effects models were applied based on the degree of heterogeneity. In contrast, Table 5 provides sensitivity analyses comparing fixed- and random-effects models for each predictor to assess the robustness of the findings. The meta-analysis identified the following factors as significantly influencing readmission after PCI: age, heart failure, diabetes, chronic lung disease, triple vessel lesion, female gender, hypertension, peripheral vascular disease, renal failure, glomerular filtration rate (GFR), N-terminal pro B-type natriuretic peptide (NT-proBNP), and triglycerides. Sensitivity analyses revealed no significant differences between fixed- and random-effects models, indicating that the pooled estimates were stable and robust.

Of the 10 studies, one reported insufficient details on model development and was excluded, leaving 9 studies eligible for further analysis. The combined AUC for these studies was 0.80 (95% CI: 0.74–0.85) (Fig. 4). However, substantial heterogeneity was observed across studies (I² = 97%, p 0.001), prompting subgroup analyses to investigate potential sources of variability in readmission following PCI (see Table 6).

We evaluated 10 risk prediction models, of which all but the study by Minges et al. [37] showing moderate to good predictive performance during internal or external validation. Reported AUC values ranged between 0.660 and 0.899. However, based on the PROBAST checklist, all studies were classified as having a high risk of bias, which limits the generalizability of these risk prediction models. The pooled AUC for the nine models included in the meta-analysis was 0.80 (95% CI: 0.74–0.85). Pooling AUC estimates across studies enables a quantitative synthesis of the discriminative performance of readmission prediction models following PCI. Although AUC is a widely used and robust metric for assessing a model’s ability to rank patient risk, it is sensitive to outcome prevalence and calibration characteristics [41]. Thus, while meta-analysis of AUCs facilitates comparison of model performance across heterogeneous settings, it should not be interpreted as a direct indicator of model generalizability. To account for potential biases arising from differences in model specifications—such as variations in predictor sets or outcome timeframes—subgroup analyses were conducted based on publication year, study region, readmission type, and modeling approach. High heterogeneity across the studies may be attributed to differences in outcome definitions or methodological approaches.

Despite variability in performance and quality, these models provide valuable insights for future research. For instance, the study by Minges et al. [37], which employed a large-sample prospective design, relied on random split validation for internal validation. Although this method is a type of internal validation, it does not account for issues such as model overfitting [42], rendering it less favorable. The study by Zhang et al. [38] faced challenges such as a sample size yielding an EPV ratio below 20 and data from a single-center retrospective design in northwest China, leading to risks of bias in the participants, predictors, and outcome domains. Nevertheless, the study excelled in the analysis domain by using multiple imputation for missing data (10%) and employing three robust predictor selection methods: LASSO regression, random forest, and best subset selection. Some models overlooked the importance of avoiding univariate analysis during predictor screening, a limitation that should be addressed in future studies.

The study by Liu et al. [32] integrated traditional logistic regression with machine learning methods during model development. Evidence suggests that, compared to traditional approaches, machine learning offers superior nonlinear fitting capabilities, enabling the capture of complex relationships and improving both predictive accuracy and model robustness [43]. Machine learning also demonstrates advantages in handling large-scale, high-dimensional, and incomplete datasets, and supports continuous model updating as new data become available—enhancing adaptability and long-term performance. However, machine learning methods are not without limitations. They are prone to overfitting, particularly when applied to small datasets; often lack transparency, reducing interpretability relative to traditional statistical models; and typically require large volumes of high-quality, complete data to function optimally. Moreover, the “black box” nature of some advanced algorithms may hinder their clinical applicability in contexts where interpretability is essential. Traditional methods such as logistic regression remain valuable, especially in scenarios involving smaller datasets or when clinical interpretability is prioritized. Ultimately, model selection should be guided by the specific characteristics of the research question, including data size and quality, the complexity of the associations involved, and the trade-off between interpretability and predictive performance.

Recent advances in machine learning have substantially improved risk prediction models in cardiovascular medicine. For example, Yilmaz et al. [44], demonstrated that machine learning algorithms leveraging electrocardiogram (ECG) features—such as P-wave, QRS complex (the combination of Q, R, and S waves in the electrocardiogram representing ventricular depolarization), and T-wave characteristics—can accurately predict obstructive coronary artery disease in patients undergoing high-risk treadmill exercise testing. Similarly, Cicek et al. [45] highlighted the potential of deep learning models for predicting short-term mortality in patients with acute pulmonary embolism. These studies illustrate a paradigm shift toward more sophisticated, data-driven approaches in cardiovascular risk prediction. While the current review primarily focused on conventional statistical models, future research should increasingly incorporate advanced machine learning techniques to enhance the accuracy of post-PCI readmission prediction. Standardized machine learning frameworks that integrate multimodal data—including clinical, imaging, and electrophysiological variables—may outperform traditional risk scores and offer greater predictive utility in diverse clinical settings.

Overall, while the included models exhibited moderate to good performance, the high risk of bias underscores the need for improvements. Future research should focus on optimizing sample sizes, addressing missing data, refining predictor selection, accounting for data complexity, and improving model fitting.

The frequently identified predictors carry significant implications for nursing practice and future research.

Elderly patients are at heightened risk of hospital readmission due to age-related declines in cognitive and organ function, diminished self-care capacity, and a higher burden of comorbidities [46, 47]. These risk factors are comprehensively captured by the Intermountain Risk Score (IMRS), which integrates age, sex, and routine laboratory markers to predict post-PCI outcomes. IMRS has demonstrated strong prognostic value for both short- and long-term mortality in patients with ST-segment elevation myocardial infarction (STEMI) and cardiogenic shock, with advanced age being a heavily weighted component [48, 49]. These findings highlight the importance of targeted management strategies for elderly patients, including enhanced postoperative monitoring and individualized rehabilitation programs.

Female sex is independently associated with higher readmission rates following STEMI, a disparity that is reflected in IMRS-based risk stratification. While biological factors—such as heightened pain sensitivity and a greater comorbidity burden—partially explain this difference [50, 51, 52, 53, 54, 55, 56, 57], the IMRS further quantifies sex-specific risk, facilitating early identification of high-risk female patients [48]. Given the validated prognostic utility of IMRS in predicting post-discharge adverse events [49], structured interventions—including motivational follow-up and psychological support—should be prioritized for this subgroup to improve outcomes.

The bidirectional relationship between diabetes and coronary heart disease (CHD) complicates the postoperative recovery of patients with coexisting conditions. Stress-related hyperglycemia during angina or myocardial infarction, combined with vascular endothelial damage from high glucose levels, exacerbates CHD [58]. Consequently, patients with diabetes face heightened readmission risks due to challenges in disease management. Clinical practices should emphasize frequent follow-ups, strict blood glucose control, and targeted interventions to slow plaque progression, thereby reducing recurrent acute coronary syndrome (ACS) and readmission rates. A comprehensive approach, considering the interplay of diabetes, cardiovascular disease, and aging can lead to better treatment outcomes and enhanced patient quality of life.

Impaired cardiac output in heart failure patients increases their vulnerability to adverse cardiovascular events following PCI, elevating the risk of readmission. Interventions should focus on managing heart failure symptoms and optimizing post-PCI care to minimize complications.

The coexistence of chronic lung disease and cardiovascular disease intensifies readmission risks due to shared inflammatory pathways and oxygen supply-demand imbalances. Studies indicate that integrated prevention strategies enhance treatment efficacy [59]. Mechanisms such as inflammatory activation leading to plaque instability and myocardial hypoxia further exacerbate cardiovascular risk [60]. Observational data show a 3.8-fold increase in adverse cardiovascular events within 30 days after acute exacerbations of chronic obstructive pulmonary disease (COPD) [61]. For PCI patients with concurrent lung disease, integrating lung rehabilitation into care plans is crucial.

Patients with multivessel disease face higher risks of MACE due to extensive myocardial ischemia [62]. Studies associate this condition with severe atherosclerosis, poorly managed comorbidities, and compromised cardiac function [63]. A staged PCI approach, treating the culprit vessel first and addressing remaining lesions later, is recommended to improve outcomes [64].

The clinical implementation of readmission prediction models for post-PCI patients remains critical. However, further clinical trials are necessary to validate their efficacy in reducing readmission rates and expanding their applicability to broader patient populations. Such efforts would enhance the utility of these models in optimizing patient outcomes.