Meta-analysis and external validation of a risk model for gastrointestinal bleeding after percutaneous coronary intervention Authors: Hualong Ma, Cong Peng, Xiaoge Liu, Yuexin Huang, Jiahui Liu, Haobin Liang, Lizhen Deng, Ying Zhang, Dalong Chen, Qiaohong Yang Published: 2025 ABSTRACT AbstractBackgroundPercutaneous coronary intervention (PCI) is a cornerstone in the management of coronary artery disease; however, postoperative gastrointestinal bleeding (GIB) represents a significant complication that adversely impacts patient prognosis. Numerous factors influence GIB, yet no comprehensive meta-analysis has synthesized these to date. Current predictive tools, such as the CRUSADE and PRECISE-DAPT scores, exhibit limited efficacy in forecasting GIB following PCI, underscoring the urgent need for a more precise model to enhance risk management.MethodsThis study employed a meta-analysis to identify risk factors for GIB post-PCI and subsequently developed predictive models based on these findings. The meta-analysis incorporated 77 studies encompassing a total of 7,211,114 patients with PCI. Ten predictive models were constructed from the analysis and validated in an external cohort of 3425 patients with PCI from two tertiary hospitals.ResultsA total of 129 influencing factors were included, with a meta-analysis conducted on 71, identifying 60 factors significantly associated with GIB. Model I, the most clinically applicable model comprising nine risk factors (female sex, advanced age, smoking, prior gastrointestinal ulcer, renal insufficiency, non-use of proton pump inhibitors, anticoagulant use, anemia, and glycoprotein IIb/IIIa receptor antagonist administration), demonstrated superior performance in external validation with an AUC of 0.842. This outperformed the CRUSADE score (AUC = 0.770) and PRECISE-DAPT score (AUC = 0.772), with DeLong’s test, significant positive Net Reclassification Improvement Index, and Integrated Discrimination Improvement further confirming its enhanced clinical utility.ConclusionsGIB following PCI is influenced by a multitude of factors. Model I excels in predicting this complication, surpassing existing scoring systems, and offers substantial clinical value by enabling personalized risk management to improve patient outcomes. All nine predictors are routinely available at the bedside or in the electronic health record, facilitating immediate clinical implementation without additional testing.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12916-025-04591-0.Keywords: Percutaneous coronary intervention, Gastrointestinal bleeding, Risk factors, Meta-analysis, Predictive model Background Percutaneous coronary intervention (PCI) is a cornerstone in the management of coronary artery disease, providing a minimally invasive approach to restore coronary blood flow [1]. Despite its efficacy, PCI is associated with significant complications, notably bleeding events [1]. Among these, gastrointestinal bleeding (GIB) is predominantly a medication-related complication, stemming from the mandatory antithrombotic therapies, such as antiplatelets and anticoagulants used to prevent post-PCI thrombosis [2]. GIB occurs in approximately 0.4% to 4.3% of patients with PCI [3], accounting for 61.7% of all bleeds [4], with its severity amplified by these pharmacologic risks. This complication is particularly severe due to its association with increased healthcare expenditures [5], prolonged hospital stays [6], and elevated short- and long-term mortality [7, 8]. The mandatory use of antiplatelet and anticoagulant therapies to prevent thrombotic complications post-PCI significantly elevates the risk of GIB, making it a critical clinical challenge that demands targeted risk assessment and management [2]. Some studies have explored potential risk factors for post-PCI GIB, including demographic factors such as age and sex, comorbidities like gastrointestinal ulcers, and medication use such as anticoagulants [9, 10]. However, the literature is fragmented, with inconsistencies arising from small sample sizes, heterogeneous patient populations, and varying definitions of GIB [11–13], limiting the generalizability of findings. Furthermore, existing predictive tools, such as the CRUSADE [14] and PRECISE-DAPT scores [15], are inadequate for this specific complication. The CRUSADE score, designed for in-hospital major bleeding in patients with non-ST-segment elevation myocardial infarction, has shown limited discriminatory power for GIB in the post-PCI setting. In a Chinese cohort study of 4,943 patients with PCI, the CRUSADE score yielded an area under the receiver operating characteristic curve (AUC) of 0.715 for predicting 1-year GIB risk [16]. Similarly, the PRECISE-DAPT score was developed to predict overall out-of-hospital major bleeding in patients on dual antiplatelet therapy after PCI [15]. A recent meta-analysis confirmed its moderate discriminative ability for major bleeding (pooled C-index ≈ 0.71), but it was not validated for GIB prediction [17]. This suboptimal performance, coupled with the absence of a dedicated tool for GIB, underscores the need for a novel scoring system specifically designed to predict GIB risk in patients with PCI. A pooled, meta-analytic synthesis of adjusted effect estimates can increase precision, reconcile inconsistent findings across settings, and identify reproducible predictors that are clinically actionable. We therefore conducted a comprehensive meta-analysis to quantify risk factors for GIB after PCI, converted pooled effect sizes into weighted predictor coefficients, and used those weights to derive a points-based risk score. Finally, we externally validated the resulting model in an independent two-hospital cohort to assess discrimination and clinical utility with the goal of supporting management of post-PCI GIB. Methods Participants Derivation cohort The protocol and analysis plan for this meta-analysis were developed a priori. The derivation cohort for this study was established through a meta-analysis of 77 original studies, in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) reporting guidelines [18]. A comprehensive literature search was performed across multiple databases, including MEDLINE, PubMed, Web of Science, CINAHL, Embase, Cochrane Library, SinoMed, China National Knowledge Infrastructure, Wanfang Database, and VIP Database. The search was unrestricted by language and encompassed the period from the inception of each database to June 21, 2025. Studies were included based on the following criteria: (1) patients with PCI; (2) studies reporting factors associated with GIB following PCI or predictive models for such events; (3) case–control studies or cohort studies; (4) studies reporting adjusted OR or RR with 95% CI for factors related to GIB, accounting for confounding variables. Exclusion criteria were as follows: (1) studies with incomplete data or where the original full-text article could not be obtained; (2) duplicate publications; (3) studies with a score of ≤ 6 on the Newcastle–Ottawa Scale [19]. Review articles were excluded from the meta-analysis; however, their references were examined to identify additional studies meeting the inclusion criteria. Randomized controlled trials were not intended for inclusion since they principally assess intervention effects, often do not report adjusted baseline associations required for this meta-analysis of influencing factors, and their selective enrollment or intervention arms could alter baseline risk relationships and introduce heterogeneity if combined with observational prognostic data. The detailed literature search strategy is provided in Additional file: section “Search Strategy.” The literature selection process is illustrated in Fig. 1A. Five authors (HLM, XGL, CP, YXH, HBL) independently screened titles and abstracts during the initial selection, followed by a full-text review for the second phase. Data from the final included studies were extracted by three authors (HLM, XGL, CP) independently and subsequently integrated. In cases of duplicate publications, the most recent version was selected. The quality of the included studies was evaluated by three authors (HLM, YXH, HBL) independently using the Newcastle–Ottawa Scale, with higher scores indicating better methodological quality. Any discrepancies during the screening, extraction, or quality assessment processes were resolved through discussion with the remaining authors. PRISMA flow diagram for meta-analysis and validation cohort. Note: Panel A represents the meta-analysis; panel B represents the validation cohort External validation cohort The external validation cohort was derived from two independent tertiary hospitals in China. A total of 8340 patients with PCI were initially enrolled, and following screening, 3425 cases were included, spanning the period from January 1, 2015, to December 31, 2024. The inclusion criteria for the external validation cohort were: (1) Age ≥ 18 years; (2) Diagnosis of coronary artery disease; (3) Patients with PCI. Exclusion criteria included: (1) Patients without access to medical care; (2) Patients with a data missing rate > 20%. The selection process for the external validation cohort is depicted in Fig. 1B. Outcome The GIB definition was pre-specified. Because included studies used different operational definitions of GIB, we created a study-level mapping of each study’s GIB definition (Additional file: section “Characteristics of the Included Studies”) and categorized definitions as: Validated (TIMI [20]/BARC [21]/GUSTO [22]), Clinical overt (hematemesis/melena/hematochezia), Lab/transfusion (hemoglobin drop or transfusion), endoscopic/imaging-confirmed, Administrative/ICD-only, or Unclear. The composite GIB outcome used for pooling considered any of the following: overt GI bleeding, Hb drop ≥ 2 g/dL attributed to GI bleeding, transfusion for suspected GI bleeding with clinical evidence, endoscopic/radiologic confirmation, or validated bleeding score designations mapped to the composite. To assess the impact of heterogeneous definitions, we performed a sensitivity analysis excluding Administrative/ICD-only and Unclear/NR studies and compared the performance of the best performed model using DeLong’s test, meanwhile calculated Net Reclassification Improvement (NRI) and Integrated Discrimination Improvement (IDI) (sensitivity analysis task 1). Severe GIB was defined by meeting at least one of the following established criteria: (1) a hemoglobin decrease of ≥ 5 g/dL (classified as thrombolysis in myocardial infarction (TIMI) life-threatening bleeding); (2) Bleeding Academic Research Consortium (BARC) type 3 or 5 bleeding (requiring surgical intervention, vasopressor support, or resulting in fatality); (3) Global Utilization of Streptokinase and Tissue Plasminogen Activator for Occluded Coronary Arteries (GUSTO) moderate or severe bleeding (overt bleeding accompanied by a significant hemoglobin decline). Exclusion criteria were: bleeding from esophagogastric varices, hemorrhoids or anal fissures, and false positives resulting from medications (e.g., iron supplements, bismuth compounds) or dietary artifacts. Statistical methods Meta-analysis For each risk factor or predictive model associated with post-PCI GIB, we extracted reported effect measures (adjusted odds ratios (ORs), hazard ratios (HRs), or risk ratios (RRs)) and their 95% confidence intervals and pooled them using inverse-variance weighting on the log scale. Given the overall pooled GIB incidence is 1.06%, which is very low, we employed the rare-event approximation (logHR ≈ logRR ≈ logOR) to harmonize measures. Where study-level baseline event rates (p₀) were available, we converted RRs to ORs using OR = [RR × (1 − p₀)] / [1 − p₀ × RR]. Standard errors for log-effect estimates were calculated from reported 95% confidence intervals (CI) using SE = (log(upper CI) − log(lower CI)) / (2 × 1.96), and inverse-variance weighting was applied on the log(OR) scale. All statistical tests were two-sided, and analyses were conducted using RStudio (Version 2024.12.1 + 563). The choice between a fixed-effect or random-effects model was guided by heterogeneity, which was assessed with the Cochrane Q test and quantified by the I2 statistic [23, 24]. A random-effects model was employed when significant heterogeneity was present, defined by a Q-test P-value < 0.10 or an I2 > 60%. For analyses including three or more studies, we performed sensitivity analyses to assess the robustness of the results (see “panel (5)” in Additional file: section “Specific Meta-Analysis”). If the exclusion of a single study during sensitivity analysis resolved significant heterogeneity, that study was assessed for being a statistical outlier; if not an outlier, it was retained, and the random-effects model was maintained, as this approach aligns with our goal of developing a generalizable GIB prediction model rather than one limited to highly homogeneous populations. We evaluated potential publication bias through the visual inspection of funnel plots and formal testing with Egger’s regression test (P < 0.05) [25]. Where bias was suspected, the trim-and-fill method was used to estimate and adjust for missing studies [26]. All statistical tests were two-sided. Model development A predictive risk score was constructed based on the significant risk factors identified in the meta-analysis. When original studies reported HRs or RRs, those measures were converted to ORs as described above prior to pooling; where conversion was not possible, the rare-event approximation was applied. To create a points-based scoring system, the OR derived from the meta-analysis for each risk factor was converted to its corresponding β-coefficient using the natural logarithm function (β = ln(pooled OR)). Subsequently, each β-coefficient was scaled by multiplying it by 10 and rounding the result to one decimal place. These transformed coefficients were then used to assign point values to risk categories defined by pre-specified thresholds, ensuring that higher-risk categories received proportionally higher points to reflect the magnitude of association (e.g., a β of 0.5 after scaling yields 5 points for a high-risk category). Thresholds for continuous variables, such as age, were determined by meta-analysis-derived quantiles in the pooled cohort (e.g., < 64 years = 0 points; 65–74 years = 1 point; 75–84 years = 2 points; ≥ 85 years = 3 points) and anchored to established clinical guidelines defining ≥ 65 years as a bleeding risk factor [27]. For categorical predictors (e.g., presence/absence of smoking), binary thresholds were applied, assigning points only to the risk-present category based on consistent reporting across studies and diagnostic criteria. This approach facilitates easy manual summation in clinical settings—a common approach to simplify risk scoring for practical use, as described by Sullivan et al. [28]. To assess the impact of the harmonization approach, we performed sensitivity analyses of the model performance converting or not RR/HR to OR (sensitivity analysis task 2). After inspecting reporting frequencies across 77 studies, and to systematically evaluate the trade-off between predictive performance and clinical parsimony, ten nested models (A–J) were developed, progressively incorporating additional predictors based on reporting frequency (≥ 10 studies) and effect size, with scoring criteria detailed in Additional file: section “Scoring Criteria.” We selected this threshold because meta-analytic diagnostics commonly used to assess between-study heterogeneity and small-study effects, for example funnel-plot asymmetry and related tests, are widely considered unreliable when applied to fewer than ~ 10 studies [29, 30]; requiring ≥ 10 studies therefore improves the stability and interpretability of pooled effect estimates that inform predictor weighting. Predictor effect sizes were pooled and converted to score weights so that the magnitude of association determined point allocation even where frequency governed initial prioritization. We adopted such methodology since this approach balances practicability (selecting predictors that are likely to be available in routine datasets) with effect-size driven weighting in the score derivation. As a sensitivity analysis task 3, to test whether selecting predictors by “strength of association,” rather than by “reporting frequency” materially changes discriminative performance, we constructed the model as follows: (1) rank all pooled ORs from highest to lowest. If the OR value is significantly less than 1, take its reciprocal, (2) select the top N predictors where N equals the number of predictors ultimately included in the best model developed via “reporting frequency” considering fair comparison, and (3) fit Model M using the same weighting procedure (log-OR → β → points). Ultimately, we compared AUCs between the two models using the DeLong test, and calculated NRI and IDI. Model validation External validation of the risk prediction model was conducted using retrospective cohort data from two independent tertiary hospitals in China to evaluate its generalizability. Little’s test indicated that missing data were not completely at random (χ2 = 1535.62, P < 0.001) [31]. To preserve the representativeness of real-world clinical scenarios while ensuring data quality, we excluded cases with more than 20% missing values [32]. For the remaining cohort, we addressed residual missingness using multiple imputation, generating 10 datasets for analysis [33]. For continuous variables, normality was assessed via Q-Q plots. Variables approximating normal distributions were described using mean ± standard deviation, while non-normal distributions were summarized with median and interquartile range. Between-group comparisons utilized independent samples t-tests for equal variance assumptions or Wilcoxon rank-sum tests for heterogeneous variances. Categorical variables were presented as counts and percentages, with group differences analyzed using Fisher’s exact test or chi-square tests, as appropriate. Model performance was evaluated using accuracy, precision, recall, F1 score, area under the receiver operating characteristic curve (AUC), and decision curve analysis. Discrimination differences between models were assessed via DeLong’s test [34]. Because no consensus, externally validated risk score exists specifically for predicting GIB after PCI, and as external comparators, we evaluated two established bleeding risk scores—CRUSADE (an in-hospital bleeding score for NSTEMI) [14] and PRECISE-DAPT (an out-of-hospital bleeding score during dual antiplatelet therapy after PCI) [15]—and summarized their key component variables for methodological context (CRUSADE: baseline hematocrit, creatinine clearance, heart rate, systolic blood pressure, female sex, signs of heart failure at presentation, diabetes, and peripheral vascular disease; PRECISE-DAPT: age, hemoglobin, white blood cell count, creatinine clearance, and prior spontaneous bleeding). Both scores were computed exactly as originally published using variables available in our dataset. All comparative performance metrics (AUC, DeLong tests, and calculating NRI and IDI) compared each score to the same composite post-PCI GIB outcome and identical follow-up window used to evaluate our model. Risk stratification and cut-point selection To select an operating threshold for binary classification, we computed Youden’s J statistic across all possible total-score thresholds in the external validation cohort and chose the threshold that maximized J as the “optimal” cut-off for high versus non-high risk [35, 36]. We then reported observed GIB rates across these strata. Ethical consideration The study utilized anonymized patient data, ensuring compliance with the ethical standards set forth in the Declaration of Helsinki. Ethical approval was granted by the Institutional Review Board of the First Affiliated Hospital of Jinan University (Approval No. JNUKY-2021–001). Given the retrospective nature of the analysis, the IRB determined that obtaining individual informed consent was unnecessary, as the data were derived from existing, anonymized records without direct patient involvement. All measures were taken to protect participant privacy and ensure the integrity of the research process. Data-sharing statement Deidentified data from this study may be requested from the corresponding author QHY (yqiaohong@163.com) or first author HLM (Ma980383402@outlook.com) for scientific research purposes, subject to ethical approval. Results Description of the cohort Meta-analysis A systematic literature search identified 11,247 articles, with 77 studies ultimately included (see Fig. 1A for the study selection flowchart) [4, 9–13, 16, 37–106]. All included studies had NOS scores > 6, indicating high methodological quality (quality assessment detailed in Additional file: section “Quality Assessment”). The pooled cohort comprised 7,211,114 cases, including 76,510 GIB-positive cases, yielding a pooled GIB incidence of 1.06%. Detailed study characteristics are provided in Additional file: section “Characteristics of the Included Studies.” We extracted individual-level baseline variables from the 77 included studies and computed population-weighted summaries. The derivation cohort had a weighted mean age of 64.6 years (range of study-level means: 46.4–73.1), with 33.1% female patients. Prevalence of key risk factors was as follows: smoking (26.4%), prior gastrointestinal ulcer (0.5%), renal failure (8.1%), anemia (8.5%), hypertension (67.1%), diabetes (31.2%), proton pump inhibitor (PPI) use (36.0%), anticoagulant use (2.8%), and IIb/IIIa antagonist use (20.2%). It should be noted that not all studies reported data for every variable; therefore, the denominators used to calculate these proportions reflect the subset of studies that provided relevant information for each specific characteristic. We identified 129 potential GIB risk factors, categorized into 11 domains: demographic factors, anthropometric measurements, compliance/habits, comorbidities, preoperative characteristics, laboratory biomarkers, medication use, periprocedural/healthcare process variables, patient condition, hospital-related variables, and genetic variables (variable coding from original studies in Additional file: section “Variables Codes”). Of these, 71 factors with data reported in ≥ 2 studies were included in meta-analyses. Results revealed 60 factors significantly associated with GIB risk. Detailed meta-analysis outcomes for each factor are summarized in Additional file: section “Specific Meta-Analysis,” with pooled results presented in Additional file: section “Summary of the Meta-Analysis,” and forest plots in Fig. 2. Forest plot of all meta-analyzed risk factors for gastrointestinal bleeding Validation cohort An external validation cohort comprising 3425 patients with PCI was established across two hospitals, with 166 cases (4.85%) experiencing GIB. Comparative analyses between GIB-positive and GIB-negative groups were detailed in Table 1. Table 1Baseline characteristics of the validation cohortCharacteristicsOverall (n = 3425)GIB-negative (n = 3259)GIB-positive (n = 166)t/FpGender Female863 (25.2)804 (24.67)59 (35.54)9.3370.002 Male2562 (74.8)2455 (75.33)107 (64.46)Smoking Yes2282 (66.63)2188 (67.14)94 (56.63)7.3820.007 No1143 (33.37)1071 (32.86)72 (43.37)Age63.04 ± 8.7662.67 ± 8.3270.26 ± 12.99−7.45< 0.001Gastrointestinal ulcer No3195 (93.28)3063 (93.99)132 (79.52)50.495< 0.001 Yes230 (6.72)196 (6.01)34 (20.48)Renal insufficiency No3057 (89.26)2935 (90.06)122 (73.49)43.481< 0.001 Yes368 (10.74)324 (9.94)44 (26.51)Anemia No2473 (72.2)2387 (73.24)86 (51.81)35.104< 0.001 Yes952 (27.8)872 (26.76)80 (48.19)Hypertension No1633 (47.68)1571 (48.2)62 (37.35)7.0330.008 Yes1792 (52.32)1688 (51.8)104 (62.65)Diabetes mellitus No2593 (75.71)2488 (76.34)105 (63.25)14.012< 0.001 Yes832 (24.29)771 (23.66)61 (36.75)Proton pump inhibitors No1175 (34.31)1082 (33.2)93 (56.02)35.504< 0.001 Yes2250 (65.69)2177 (66.8)73 (43.98)Anticoagulants No2523 (73.66)2449 (75.15)74 (44.58)74.508< 0.001 Yes902 (26.34)810 (24.85)92 (55.42)IIb/IIIa receptor antagonist No2759 (80.55)2678 (82.17)81 (48.8)110.218< 0.001 Yes666 (19.45)581 (17.83)85 (51.2) Baseline characteristics of the validation cohort The model Model development Ten models (A–J) were developed, with scoring criteria listed in Additional file: section “Scoring Criteria.” Predictor weights were assigned according to pooled ORs, detailed in Table 2. For clinical utility, age per 1-year increment was transformed into ORs per 10-year increment. Table 2Pre-defined models and their performanceModelNumber thresholdIncluded factorsAccuracyPrecisionRecallF1AUCA≥ 40Gastrointestinal ulcer0.9040.1480.2050.1720.572B≥ 28Gastrointestinal ulcer, renal insufficiency0.8250.1220.4220.1900.638C≥ 22Gastrointestinal ulcer, renal insufficiency, proton pump inhibitors0.5710.0810.7590.1460.703D≥ 19Age, gastrointestinal ulcer, renal insufficiency, proton pump inhibitors0.5320.4820.8430.6130.752E≥ 18Age, smoking status, gastrointestinal ulcer, renal insufficiency, proton pump inhibitors0.7760.6290.6330.6310.752F≥ 17Gender, age, smoking status, gastrointestinal ulcer, renal insufficiency, proton pump inhibitors0.7160.6120.6990.6530.757G≥ 14Gender, age, smoking status, gastrointestinal ulcer, renal insufficiency, proton pump inhibitors, anticoagulants0.7580.6320.7170.6720.792H≥ 13Gender, age, smoking status, gastrointestinal ulcer, renal insufficiency, proton pump inhibitors, anticoagulants, anemia0.6980.6360.7950.7070.810I≥ 11Gender, age, smoking status, gastrointestinal ulcer, renal insufficiency, proton pump inhibitors, anticoagulants, anemia, IIb/IIIa receptor antagonist0.8010.7560.7590.7570.842J≥ 10Gender, age, smoking status, gastrointestinal ulcer, renal insufficiency, proton pump inhibitors, anticoagulants, anemia, IIb/IIIa receptor antagonist, hypertension, diabetes mellitus0.8300.7830.7230.7520.844K (sensitivity analysis task 1)≥ 11Gender, age, smoking status, gastrointestinal ulcer, renal insufficiency, proton pump inhibitors, anticoagulants, anemia, IIb/IIIa receptor antagonist0.7820.7780.7530.7650.835L (sensitivity analysis task 2)≥ 11Gender, age, smoking status, gastrointestinal ulcer, renal insufficiency, proton pump inhibitors, anticoagulants, anemia, IIb/IIIa receptor antagonist0.7870.7720.7830.7770.846M (sensitivity analysis task 3)OR ≥ 3.57Albumin, intraoperative hypotension, creatinine clearance, inotropic agents, heparin + IIb/IIIa receptor antagonist during surgery, emergency surgery, NYHA class, rectum/anal cancer, statins0.7160.6800.6690.6740.738 Pre-defined models and their performance For continuous variables in our pre-specified models, cutoffs of age (< 64 years = 0 points; 65–74 years = 1 point; 75–84 years = 2 points; ≥ 85 years = 3 points) were meta-analysis-driven in the pooled cohort but meanwhile anchored to the guidelines, which define ≥ 65 years as a risk factor [27]. Binary thresholds for categorical predictors, such as whether smoking or not, reflected consistent reporting across studies and established diagnostic facts, namely presence/absence of these conditions as binary risk modifiers. Model validation Independent external validation of Models A–J, CRUSADE, and PRECISE-DAPT scores was conducted using the two-hospital cohort. ROC, which demonstrates a model’s discriminative ability by showing how well it distinguishes between patients who will and will not experience GIB, and decision curves, which assess the net clinical benefit across a range of risk thresholds, were presented in Fig. 3A and B, respectively. Performance metrics (accuracy, precision, recall, F1, AUC) were summarized in Table 2. A trend of increasing AUC and improved decision curve positioning toward the top-right quadrant was observed with greater variable inclusion. However, DeLong’s test revealed no statistically significant difference between Models I and J (Z = − 1.242, P = 0.428). Model I was selected as the optimal model over Model J because, considering clinical usability and the need for a practical, parsimonious instrument, Model I achieved comparable performance metrics with fewer variables, making it more suitable for bedside application. Performance of the developed gastrointestinal bleeding risk prediction model. Note: Panel A represents ROC curves comparing Models A–J with CRUSADE and PRECISE-DAPT in the validation cohort; panel B represents decision curve analysis of net clinical benefit; panel C represents predicted probability of GIB across Model I scores; panel D represents calibration plot for Model I with Hosmer–Lemeshow test; panel E represents observed GIB incidence by Model I risk strata (low, medium, high). Panel F represents the interface of the web calculator powered by Model I Model I significantly outperformed both CRUSADE (Z = − 32.242, P < 0.001) and PRECISE-DAPT (Z = − 27.843, P < 0.001). Besides, as it was shown in Table 3, when compared with the CRUSADE score, Model I demonstrated a significant IDI of 0.075 (P < 0.001), indicating a substantial overall gain in the separation of event and non-event probabilities. The NRI was 0.078 and was also statistically significant (P = 0.032), suggesting that Model I correctly moved 7.8% more patients to appropriate risk categories (upward for GIB-positive cases and downward for GIB-negative cases) compared to the CRUSADE score. Similarly, when compared to the PRECISE-DAPT score, Model I showed an IDI of 0.089 (P = 0.034), again confirming its enhanced discriminatory power. The NRI was 0.069 (P = 0.018), meaning Model I correctly reclassified 6.9% more patients than the PRECISE-DAPT score. These findings demonstrate that Model I provides a statistically superior and clinically meaningful improvement in risk stratification for post-PCI GIB compared to existing general bleeding scores. Table 3Model performance comparison resultsComparisonNRIP for NRIIDIP for IDIΔAUCP for DeLong testModel I vs CRUSADE0.0780.0320.075< 0.0010.072< 0.001Model I vs PRECISE-DAPT0.0690.0180.0890.0340.070< 0.001Model K vs Model I− 0.0020.271− 0.0030.181− 0.0070.092Model L vs Model I0.0040.5810.0020.0870.0020.161Model M vs Model I− 0.102< 0.001− 0.0830.008− 0.104< 0.001CRUSADE (in-hospital major bleeding; NSTEMI): baseline hematocrit, creatinine clearance, heart rate, systolic blood pressure, female sex, signs of heart failure at presentation, diabetes mellitus, and history of peripheral vascular diseasePRECISE-DAPT (out-of-hospital bleeding during DAPT after PCI): age, hemoglobin, white blood cell count, creatinine clearance, and prior spontaneous bleeding Model performance comparison results CRUSADE (in-hospital major bleeding; NSTEMI): baseline hematocrit, creatinine clearance, heart rate, systolic blood pressure, female sex, signs of heart failure at presentation, diabetes mellitus, and history of peripheral vascular disease PRECISE-DAPT (out-of-hospital bleeding during DAPT after PCI): age, hemoglobin, white blood cell count, creatinine clearance, and prior spontaneous bleeding The relationship between total Model I score and GIB positivity probability is illustrated in Fig. 3C. Calibration curves (Fig. 3D) demonstrated strong alignment with observed outcomes, confirmed by non-significant Hosmer–Lemeshow goodness-of-fit testing (χ2 = 2.324, P = 0.182). The non-significant P-value of the Hosmer–Lemeshow test indicates that there is no statistically significant difference between the predicted probabilities and the observed event rates, confirming good agreement between risk predicted by the model and the true observed risk for Model I. The results of the three sensitivity analysis tasks were presented in Table 3. ROC curves were shown in Fig. 3A. For task 1, after excluding ICD-only and unclear GIB definition studies, model discrimination (AUC) changed from 0.842 of Model I to 0.835 of Model K (Z = 1.683, p = 0.092), while the changes in both NRI and IDI were not statistically significant (NRI = − 0.002, IDI = − 0.003), indicating that the derived model’s main findings are robust to outcome definition. Meanwhile, because different studies reported different effect measures, we performed sensitivity analyses with and without converting RR/HR to OR (n = 10). For task 2, sensitivity analysis also showed that AUC changed from 0.842 of Model I to 0.846 of Model L (Z = 1.402, p = 0.161), while the changes in both NRI and IDI were not statistically significant (NRI = 0.004, IDI = 0.002), supporting the robustness of the score to the harmonization approach. For sensitivity analysis task 3, Model M performed significantly worse than Model I, with AUC differing significantly by 0.104 (Z = 3.946, p < 0.001), and both NRI and IDI were significantly less than zero (NRI = − 0.102, IDI = − 0.083), indicating that the model constructed using the “reporting frequency” method was superior to that constructed using the “strength of association (OR value)” method. Model I was selected as the primary model and was subsequently used to stratify 3,425 patients into three risk groups using tertile cut-points (low ≤ − 4.0; medium − 4.0 to 6.0; high > 6.0) based on Youden’s J. Observed GIB rates were 11/1143 (0.96%) in the low group, 19/1162 (1.64%) in the medium group, and 136/1120 (12.14%) in the high group. The difference in GIB incidence across the three groups was highly significant (χ2 = 192.66, df = 2, p < 0.001), indicating strong discrimination of Model I risk strata for actual bleeding events (Fig. 3E). A complete, stand-alone scoring method for Model I was detailed in Table 4. Table 4Points-based bedside score of model IPredictorOperational definition (use at index PCI unless stated)PointsAge< 64 years065–74 years175–84 years2≥ 85 years3SexMale0Female− 4Smoking (current)Active cigarette use within 30 days before PCI (or documented “current smoker”)7.5Prior gastrointestinal ulcerDocumented peptic ulcer disease (gastric/duodenal) by history or endoscopy within the prior 12 months or active at admission12Renal insufficiencyChronic kidney disease with eGFR < 60 mL/min/1.73 m2 (CKD ≥ stage 3) or chronic dialysis9Proton-pump inhibitor (PPI) useActive PPI therapy during the index hospitalization and/or prescribed at discharge− 10Anticoagulant therapyConcomitant oral anticoagulant (warfarin or DOAC) at discharge or planned for ≥ 7 days post-PCI (exclude procedural heparin bolus)9AnemiaBaseline hemoglobin < 13 g/dL (men) or < 12 g/dL (women)7GPIIb/IIIa receptor antagonist during PCIUse of abciximab, eptifibatide, or tirofiban periprocedurally9Total scoreSum of points above—(1) Definitions reflect pragmatic and guideline-consistent operationalizations, aligned with external validation; (2) “PPI use” carries negative points because it is protective; if no PPI is used, assign 0 Points-based bedside score of model I (1) Definitions reflect pragmatic and guideline-consistent operationalizations, aligned with external validation; (2) “PPI use” carries negative points because it is protective; if no PPI is used, assign 0 Web-based risk calculator We developed a web-based calculator (interface shown in Fig. 3F; publicly accessible at https://postpcigibrisk.pages.dev/) that initially presents a mandatory clinical disclaimer modal requiring user acknowledgment of the tool’s limitations and intended use before accessing the interface. Users then input patient-specific variables through intuitive segmented controls for demographics and interactive switches for clinical factors. The calculator dynamically displays the computed risk score via an animated gauge chart, categorizes risk level, and shows the corresponding predicted GIB incidence probability. Discussion Overview of key findings and comparison to prior work In this meta‑analysis of 77 studies, we identified a broad set of factors associated with GIB after PCI. We then used the pooled effect sizes to derive and externally validate a parsimonious points‑based risk score (Model I). Model I—composed of nine clinically available predictors: female sex, advanced age, smoking, gastrointestinal ulcer, renal insufficiency, non-use of proton pump inhibitors, anticoagulant use, anemia, and administration of IIb/IIIa receptor antagonists—showed good calibration and superior discrimination compared with commonly used bleeding instruments in an independent two-hospital validation cohort. Compared with meta-analysis of Wang L. et al. [3] and Feng et al. [107], our work addresses several important limitations. Those earlier meta-analyses were built on smaller or earlier evidence windows, and reported pooled associations without translating them into a clinically usable, externally validated risk score—factors that restrict their direct applicability to bedside risk stratification. By pooling a larger, more recent evidence base, harmonizing outcome definitions where possible, and converting pooled effects into a parsimonious points-based model with external validation, our study was designed specifically to overcome these gaps and improve clinical utility. Identified risk factors for post-PCI GIB Women face higher bleeding risks during antiplatelet therapy due to sex-specific physiological factors. Estrogen exerts antiplatelet effects [108]. Fluctuations in estrogen levels impair platelet function and coagulation mechanisms. Women also metabolize certain antiplatelet drugs more slowly [11]. This results in higher drug concentrations and amplified antiplatelet activity. A recent review confirms women have an excess bleeding risk after PCI, driven by higher incidence rates [109]. Advanced age increases GIB risk through organ dysfunction and frailty. Reduced renal and hepatic efficiency impairs drug clearance. This promotes drug accumulation and bleeding. Elderly patients often have multiple comorbidities and polypharmacy, which further heighten vulnerability [110]. Age-related gastric mucosal changes contribute as well. These include glandular atrophy, weakened defenses, such as reduced bicarbonate and prostaglandin secretion, and impaired repair [111]. Patients over 75 years are especially susceptible to injurious agents like aspirin. Recent criteria classify age ≥ 75 as a high-bleeding-risk factor, linked to hemostatic imbalance and therapy interruptions [112]. Smoking is a confirmed GIB risk factor post-PCI [11, 113]. Tobacco free radicals and toxins damage vascular walls. Nicotine induces endothelial dysfunction by inhibiting nitric oxide, raising fragility. Smoking stimulates gastric acid secretion and reduces mucosal blood flow. It also disrupts coagulation, such as by elevating fibrinogen. A 2025 analysis reports an OR of 3.22 for smoking history in GIB after PCI with DAPT; it diminishes mucosal protection and alters antiplatelet metabolism [11]. Our meta-analysis shows a dose–response relationship, with HR 2.67 for > 20 cigarettes/day. Thus, long-term smoking cessation could substantially mitigate risk [113]. Previous gastrointestinal ulcers create vulnerable mucosal sites, which may be exacerbated by antithrombotic therapy during PCI through impaired coagulation, thereby increasing GIB risk [114]. Renal insufficiency prolongs bleeding risks due to reduced clearance of anticoagulants and antiplatelet agents, leading to drug accumulation and heightened bleeding tendencies; uremic toxins and platelet dysfunction associated with kidney disease further amplify this risk [115]. Proton pump inhibitors protect the gastrointestinal mucosa by suppressing gastric acid secretion, reducing ulcer or bleeding incidence; their absence in patients with PCI undergoing antiplatelet therapy leaves the mucosa more susceptible to injury, as these drugs may impair healing and exacerbate acid-mediated damage [116]. Anticoagulants disrupt thrombus formation, thereby increasing bleeding risks such as GIB. This effect is particularly pronounced in patients with PCI, who often receive concurrent antiplatelet therapy, creating a dual challenge to normal hemostatic function [84]. Pre-existing anemia may reflect underlying conditions like occult gastrointestinal bleeding or coagulopathy, which are exacerbated by antithrombotic therapy; reduced tissue oxygenation in anemic patients may also delay mucosal healing, further elevating bleeding risks [117]. Glycoprotein IIb/IIIa inhibitors bind the platelet integrin αIIbβ3 and prevent fibrinogen and von Willebrand factor from bridging activated platelets, thereby blocking the final common step of platelet-platelet aggregation [118]. This potent terminal blockade underlies the increased bleeding risk and is modulated by agent pharmacodynamics such as “abciximab vs eptifibatide/tirofiban” [119] and by concomitant antithrombotics or patient factors such as advanced age and renal impairment [120]; these agents are frequently used in high-risk PCI cases, where their potent anticoagulant effects may amplify baseline GIB susceptibility [121]. A critical modifiable factor amplifying these risks is dual antiplatelet therapy (DAPT), which is standard post-PCI to prevent ischemic events but substantially elevates GIB incidence through combined inhibition of platelet aggregation and mucosal protection deficits. In real-world cohorts, DAPT regimens that include potent P2Y12 inhibitors such as ticagrelor or prasugrel are associated with 47% to 77% higher hazard ratios for bleeding—particularly GIB—compared with clopidogrel-based therapy [122]. GIB represents a notable proportion of hospital-recorded events, with an increase of up to twofold observed in subsets of patients with ST-elevation myocardial infarction [122]. Recent registry data from Japan further quantify this burden, reporting a 1.4% 2-year GIB hospitalization rate post-PCI, predominantly lower GIB, where DAPT use exceeded 75% and synergized with oral anticoagulant co-administration as the strongest predictor [84]. Across UK and Japanese populations, overall bleeding incidence with DAPT post-coronary intervention reaches 9–10%, underscoring the need for PPI prophylaxis—echoing Model I’s emphasis on non-use—and personalized de-escalation based on GIB risk strata [123]. This DAPT-GIB interplay highlights opportunities for targeted interventions, such as shorter durations or less potent agents in high-risk Model I profiles, to optimize net clinical benefit. Performance and comparative advantages of model I Although Model J showed comparable AUC performance to Model I with no statistically significant difference by DeLong’s test, Model I was selected for its clinical applicability. It achieved high accuracy with fewer variables, making it more suitable for bedside application. Model I demonstrated superior external validation performance compared to CRUSADE and PRECISE-DAPT scores, confirmed by higher AUC values and DeLong’s test results. These predictors differ from the existing bleeding scores in two important ways. First, gastrointestinal-specific predictors are present in Model I but absent or under-represented in general bleeding tools. GI ulcer history and PPI non-use are direct mucosal-vulnerability and protective-therapy variables respectively; both are strong, biologically plausible determinants of GIB since they represent mucosal defect + acid exposure. Because general bleeding scores, for example, CRUSADE [14], PRECISE-DAPT [15], ACUITY score [124], were developed to predict major bleeding overall rather than GIB specifically, they do not include prior peptic ulcer disease or PPI status, which is an omission that limits their sensitivity for GIB. SIGE score does include GI-specific items but focuses on re-bleeding after an index GIB rather than first-time post-PCI GIB, so its target and population differ [9]. Second, overlap on shared physiologic drivers of bleeding increases concordance across scores. Age, anemia (or hematocrit/hemoglobin), and renal dysfunction are present in Model I and in PRECISE-DAPT [15], CRUSADE [14], and ACUITY [124]; these predictors drive most of each model’s discrimination because they reflect frailty, impaired hemostasis/drug clearance, and comorbidity burden. The shared presence of these variables explains why general bleeding scores retain some yet not optimal discriminatory power for GIB. Inclusion of therapy-specific variables, for example Model I’s anticoagulant and IIb/IIIa use; ACUITY’s anticoagulant regimen term, captures treatment-related bleeding risk but may emphasize early/procedural bleeding more than delayed mucosal GIB [124]. Because Model I explicitly combines both (a) systemic hemostatic vulnerability: age, anemia, renal insufficiency, female sex, and (b) GI-specific vulnerability and prevention factors: recent ulcer history and PPI non-use, it is expected to achieve superior discrimination for post-PCI GIB specifically, which is what we observed in the external validation. However, PRECISE-DAPT [15] and CRUSADE [14] are widely embedded in clinical workflows and require fewer or more routinely captured variables, for example, bedside vitals or labs. Model I’s inclusion of prior GI ulcer and PPI use improves specificity for GIB but relies on accurate medication and GI history capture, which may not be present in administrative datasets or some registries. When rapid bedside decisions are needed and detailed GI history is unavailable, simpler scores retain practical value. Sensitivity analyses In this study, we performed three sensitivity analyses to evaluate the robustness of the developed model: task 1 excluding study that defined GIB by Administrative/ICD-only codes or where the GIB definition was unclear; task 2 assessing the impact of converting reported HRs/RRs into ORs; and task 3 comparing models with predictors selected by “reporting frequency” and “strength of association.” Results from three tasks confirmed the robustness of our model. The first result may be because, when converting β coefficients into point scores we rounded the scores, which likely attenuated some differences. The second result may be explained by (a) the low incidence of GIB, in which case the rare-event approximation (logHR ≈ logRR ≈ logOR) is appropriate, and (b) relatively few studies reported HR/RR compared with OR (n = 10 vs. 67), and not all predictors used to develop models required “HR/RR to OR” conversion during pooling. The third result suggests that prioritizing predictors with broader representation across the evidence base yields more stable and generalizable signal: frequently reported predictors are likely to be measured more consistently across settings, less affected by idiosyncratic study design, and therefore less prone to overfitting than a set chosen solely for the largest pooled effects. However, predictors with very large pooled ORs can be unstable or imprecise when they are estimated from few studies (in this study, n = 2–6), making them more susceptible to chance findings or effect-size exaggeration. Thus, the result underscores that effect size alone may overweight context-specific or sparsely measured predictors; balancing reporting frequency (availability/robustness) with magnitude of association can produce models with better discriminative performance in heterogeneous, real-world data. Practical recommendations for clinicians and researchers A key consideration for future application is whether the inputs required by Model I are commonly collected in routine practice. All nine predictors map to data elements that are standard in pre‑PCI assessment, catheterization‑lab documentation, or discharge medication reconciliation, and none requires specialized testing, imaging, or genotyping. Eight of the nine inputs (all except IIb/IIIa use) are available before or at the time of PCI, enabling pre‑emptive risk estimation and prophylaxis planning, for example, PPI initiation or antithrombotic de‑escalation. IIb/IIIa exposure is known immediately after the procedure and can update early post‑PCI risk. No new workflows or tests are required. This design choice likely explains the model’s strong discrimination alongside high usability in typical cardiology settings. We recommend using Model I when the clinical question is specifically post-PCI GIB risk. Model I gives more accurate GIB risk estimates. When rapid triage is required, established instruments such as PRECISE-DAPT or CRUSADE remain useful for approximating bleeding risk, especially early or procedural risk, since they use variables that are more consistently recorded in registries and emergency settings. Meanwhile, the SIGE score may be complementary. For patients with a recent index GIB or known labile INR, SIGE focuses on re-bleeding risk and may help decide endoscopic surveillance/secondary prevention strategies; however, SIGE is targeted and not a substitute for a first-event post-PCI GIB tool. Limitations and future directions Limitations of this study include the following: (1) Regarding the meta-analysis of risk factors, we acknowledge that identifying factors based on a low number of research source data constitutes a limitation. We addressed this influence by prioritizing variable stability over magnitude in our model development. We imposed an inclusion threshold requiring predictors to be reported in at least 10 studies to be eligible for Model I. Furthermore, our Sensitivity Analysis Task 3 comparing Model I vs. Model M empirically demonstrated that a model derived from frequently reported factors significantly outperformed a model based on high-magnitude but sparsely reported factors, confirming that our selection strategy effectively mitigated the noise associated with limited source data; (2) While our sensitivity analysis excluding administrative/ICD-only and unclear studies indicated that the derived model’s main findings are robust to outcome definition, heterogeneous GIB definitions across meta-analysis studies may affect result generalizability. Therefore, local validation and potential recalibration of the model are recommended; (3) The validation cohort derived from two tertiary hospitals in China may not represent global PCI populations. The relatively high GIB incidence in the validation cohort (4.85%) suggests potential selection bias or enhanced detection sensitivity; (4) CRUSADE and PRECISE-DAPT were not developed specifically to predict post-PCI gastrointestinal bleeding; this constraint may bias comparisons and limits direct interpretability. We chose these scores as pragmatic, real-world benchmarks because they are commonly used to guide bleeding-risk assessment in cardiology practice and because most of their component predictors are relevant to GIB risk. As antithrombotic practice and PCI techniques continue to evolve, this score should be maintained as a ‘living’ model with periodic recalibration and extension across diverse regions and care settings. Future iterations will examine time‑updated risk, incorporate treatment intensity (e.g., specific P2Y12 inhibitor, concomitant anticoagulation), access site and gastroprotective strategies, and embed web calculator or EHR‑integrated decision support. Prospective impact and fairness evaluations across clinically relevant subgroups are also planned. Conclusions In this study, we conducted a meta-analysis of risk factors for GIB following PCI and developed Model I based on these findings. Model I demonstrated superior performance in predicting post-PCI GIB, outperforming two widely used scoring systems: CRUSADE and PRECISE-DAPT. Integrating this model into clinical practice could enable personalized risk management strategies tailored to individual risk profiles, but validation in other patient cohorts is required. Supplementary Information Additional file 1: The additional file accompanying this manuscript provides a comprehensive set of supplementary data relevant to the study: Search Strategy: This section outlines the detailed search strategy employed to identify relevant studies for the meta-analysis. Quality Assessment: This section presents the quality assessment of the included studies using the Newcastle-Ottawa Scale (NOS). Variables Coding: This section provides a list of the coded risk factors and their classifications used in the meta-analysis. Characteristics of the Included Studies: This section summarizes the key characteristics of the 77 studies included in the meta-analysis, such as sample size and GIB definitions. Specific Meta-Analysis: This section presents the meta-analysis results for various risk factors associated with GIB following PCI.’ Summary of the Meta-Analysis: This section offers a concise summary of the pooled results from the meta-analysis. Scoring Criteria: This section outlines the criteria used to assign points to risk factors in the derived predictive model. Acknowledgements We would like to thank the librarians of Jinan University who helped us check the search strategy and the statisticians who helped us check all the statistics. Abbreviations Area under the receiver operating characteristic curve Bleeding Academic Research Consortium Gastrointestinal bleeding Global Utilization of Streptokinase and Tissue Plasminogen Activator for Occluded Coronary Arteries Integrated Discrimination Improvement Missing completely at random Net Reclassification Improvement Thrombolysis in myocardial infarction Authors’ contributions HLM, CP, and XGL contributed equally to this work. HLM conceptualized the study, curated the de-identified EHR dataset, performed statistical analyses, and secured funding. CP designed the methodology, validated the data extraction pipeline, and created visualization. XGL conducted literature review, drafted the initial manuscript, and interpreted clinical implications. YH, JHL, and HBL acquired institutional data resources and verified EHR variable definitions. LZD and YZ provided clinical expertise in patient data interpretation and revised critical intellectual content. DC supervised the statistical framework and validated analytical reproducibility. QHY (corresponding author) secured funding, oversaw protocol development, finalized manuscript revisions, and is accountable for all aspects of the work. All authors read and approved the final manuscript. Funding This work was supported by Guangdong Provincial Department of Education (2022KTSCX008), the Excellent Graduate Student Cultivation Program of Jinan University (2025CXY316), and the Guangdong Basic and Applied Basic Research Foundation. Data availability The datasets used and/or analyzed during the current study are available from the corresponding author Qiaohong Yang or first author Hualong Ma on reasonable request. Declarations Ethics approval and consent to participate Ethical approval was granted by the Institutional Review Board of the First Affiliated Hospital of Jinan University (Approval No. JNUKY-2021–001). Given the retrospective nature of the analysis, the IRB determined that obtaining individual informed consent was unnecessary. Consent for publication This manuscript contains no individual person’s data, including personal details, images, or videos. All analyses were performed using fully de-identified electronic health record data extracted retrospectively under IRB. The IRB granted a waiver of informed consent and authorization for publication, as the dataset contains no protected health information. Therefore, consent for publication is not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Hualong Ma, Cong Peng, and Xiaoge Liu are equal contributors. References Associated Data Supplementary Materials Additional file 1: The additional file accompanying this manuscript provides a comprehensive set of supplementary data relevant to the study: Search Strategy: This section outlines the detailed search strategy employed to identify relevant studies for the meta-analysis. Quality Assessment: This section presents the quality assessment of the included studies using the Newcastle-Ottawa Scale (NOS). Variables Coding: This section provides a list of the coded risk factors and their classifications used in the meta-analysis. Characteristics of the Included Studies: This section summarizes the key characteristics of the 77 studies included in the meta-analysis, such as sample size and GIB definitions. Specific Meta-Analysis: This section presents the meta-analysis results for various risk factors associated with GIB following PCI.’ Summary of the Meta-Analysis: This section offers a concise summary of the pooled results from the meta-analysis. Scoring Criteria: This section outlines the criteria used to assign points to risk factors in the derived predictive model. Data Availability Statement The datasets used and/or analyzed during the current study are available from the corresponding author Qiaohong Yang or first author Hualong Ma on reasonable request.