Systematic review and meta - analysis of risk prediction models for heart failure after PCI in patients with acute myocardial infarction

TL;DR

This study reviews and evaluates heart failure risk prediction models for patients with heart attacks after a common heart procedure, finding strong predictive power but high bias risks.

Contribution

The paper provides a systematic review and meta-analysis of heart failure risk models post-PCI in AMI patients, identifying key predictors and methodological shortcomings.

Findings

14 studies with 14 models were analyzed, showing strong predictive ability with AUCs between 0.847 and 0.966.

19 effective predictors of heart failure were identified, including age, biomarkers, and clinical indicators.

All studies had high bias risk due to small sample sizes and methodological issues like overreliance on univariate analysis.

Abstract

The incidence of heart failure (HF) following percutaneous coronary intervention (PCI) in patients with acute myocardial infarction (AMI) remains relatively high, severely impairing long-term prognosis and quality of life. In recent years, advances in biomarker identification and imaging technologies have driven growing research into developing HF risk prediction models for AMI patients post-PCI. However, significant heterogeneity exists across current studies in terms of model construction methods, variable selection, and validation strategies, and the predictive performance and clinical utility of these models have not been systematically evaluated.

To systematically review published studies on risk prediction models for HF after PCI in AMI patients.

Systematic review and meta-analysis.

Databases including PubMed, Web of Science, Embase, Cochrane Library, CNKI, VIP, Wanfang Data Knowledge Service Platform, Chinese Journal Full-text Database, and Chinese Biomedical Literature Database were searched from inception to December 31, 2024. Eligibility criteria included: study participants aged ≥ 18 years with AMI post-PCI; studies focused on HF risk prediction model development; and publication in Chinese or English. The Prediction Model Risk of Bias Assessment Tool (PROBAST) was used to evaluate bias risk in model construction and validation, and RevMan 5.4 software was employed for meta-analysis.

A total of 14 studies involving 14 HF risk prediction models for AMI patients post-PCI were included. All studies adopted a retrospective design, with HF incidence ranging from 3.0% to 37.5%. The area under the curve (AUC) for model discriminative ability ranged from 0.847 to 0.966, indicating strong predictive power. Meta-analysis identified 19 effective predictors of post-PCI HF in AMI patients (all P < 0.05): Killip classification, age, Gensini score, D-dimer level, N-terminal pro-brain natriuretic peptide (NT-proBNP), neutrophil-to-high-density lipoprotein cholesterol ratio (NHR), neutrophil-to-lymphocyte ratio (NLR), modified shock index (MSI), troponin I, cardiac troponin T, serum creatinine, high-sensitivity C-reactive protein (hs-CRP), ventricular wall motion amplitude, hypertension, time from symptom onset to hospital admission, number of diseased coronary vessels, diabetes mellitus, arrhythmia, and cardiac structural changes. However, all studies were rated as high overall bias risk via PROBAST, primarily due to insufficient sample size, improper handling of continuous variables, and overreliance on univariate analysis for variable selection.

While existing HF risk prediction models for AMI patients post-PCI demonstrate favorable discriminative ability, they suffer from high overall bias risk. Future studies should adhere strictly to PROBAST guidelines, optimize study design, improve methodological quality, and conduct multicenter prospective studies to validate model stability and generalizability. This will facilitate the development of clinically robust, applicable, and evidence-based prediction tools for clinical practice.

CRD42025639512

The online version contains supplementary material available at 10.1186/s12872-025-05406-z.

Full text

Acute myocardial infarction (AMI) represents the most severe form of coronary atherosclerotic heart disease, characterized by myocardial ischemia, hypoxia, and necrosis resulting from acute coronary occlusion [1]. As a leading global cause of mortality and disability, AMI imposes a particularly heavy disease burden in China [2]. In recent years, the widespread adoption of percutaneous coronary intervention (PCI), establishment of chest pain center systems, and optimization of pharmacotherapeutic regimens have significantly reduced short-term mortality in AMI patients [3]. Nevertheless, post-PCI patients remain at high risk of heart failure (HF)—a complication that not only compromises long-term prognosis and quality of life but also increases medical resource consumption and economic burden [4].

Epidemiological data underscore the clinical significance of post-PCI HF in AMI patients. For instance, a multicenter cohort study by Li et al. across 50 + hospitals in China reported that over 6.3% of AMI patients required unplanned readmission within 30 days post-PCI, with 16.7% of these readmissions directly attributed to HF [5]. This highlights HF as one of the most common post-PCI complications in AMI patients, with a direct impact on clinical outcomes. Thus, effective prediction and management of post-PCI HF risk have become urgent priorities in cardiovascular medicine.

Current clinical risk assessment tools—such as Killip classification and Gensini score—are widely used to evaluate AMI severity and coronary lesion extent. Killip classification, based on clinical symptoms, effectively reflects cardiac function [6, 7], while Gensini score indirectly quantifies myocardial injury by assessing coronary lesion severity [8, 9]. However, these tools lack specificity for post-PCI HF and may fail to accurately predict risk in distinct patient subgroups. This gap has driven research into developing HF-specific risk prediction models for AMI patients post-PCI.

Recent innovations in biomarker identification and imaging have enabled the integration of multiple predictors (e.g., age, medical history, laboratory indices like NT-proBNP and hs-CRP, and cardiac function parameters such as left ventricular ejection fraction [LVEF]) into prediction models [10]. Despite this progress, heterogeneity in model development, variable selection, and validation limits the generalization of findings. To address this, the present study systematically reviews and meta-analyzes existing models to evaluate their methodological quality, predictive performance, and clinical applicability—ultimately providing evidence for clinical practice and guiding future model optimization.

Materials and methods

Protocol and registration

This study was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines and registered on PROSPERO (Registration No.: CRD42025639512).

Inclusion and exclusion criteria

Inclusion criteria: (1) The data are incomplete or invalid, and the full text cannot be obtained; (2) Study focus: Development of HF risk prediction models; (3)Study design: Case-control, cohort, or cross-sectional studies; (4)Language: Chinese or English.

Exclusion criteria: (1) Incomplete/invalid data or unavailable full text; (2) Insufficient description of model construction methods; (3) Duplicate publications; (4) No formal prediction model development or missing model data; (5) HF listed as one of multiple outcomes (not the primary focus); (6) Reviews, case reports, conference abstracts, or non-original studies.

Literature search strategy

Publicly published studies on the construction of risk prediction models for HF after PCI in AMI patients were searched in PubMed, Web of Science, Embase, Cochrane Library, CNKI, VIP, Wanfang Data Knowledge Service Platform, Chinese Journal Full-text Database, and Chinese Biomedical Literature Database. The search period was from the establishment of each database to December 31, 2024. A combination of subject terms and free terms was used for searching, adjusted according to the characteristics of each database, supplemented by reference backtracking.

Chinese search terms: Acute myocardial infarction/acute myocardial infarction, myocardial infarction, percutaneous coronary intervention/PCI, heart failure/heart failure, model, prediction model, risk prediction, predictor, risk stratification, risk score, risk assessment;

English search terms: Myocardial infarction, percutaneous coronary intervention, PCI, heart failure, predict, prediction model, risk prediction, risk assessment, risk evaluation, risk score, risk stratification model, prediction tool, prognostic model, stratification model, post-PCI.

Literature screening and data extraction

Two independent reviewers screened titles/abstracts against eligibility criteria, followed by full-text review for final inclusion. A data extraction form was pre-designed using the Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies (CHARMS) [11]. Discrepancies were resolved via discussion, with a third reviewer consulted if consensus was not reached.

Quality assessment

The PROBAST tool was used to evaluate bias risk (4 domains, 20 items) and applicability (3 domains) [12]. A study was rated as high overall bias risk if any bias domain was classified as high or unclear; high overall applicability risk was assigned if any applicability domain was high. Two reviewers independently conducted assessments, with discrepancies resolved via discussion or third-party consultation.

Statistical analysis

RevMan 5.4 was used for meta-analysis of common predictors across models. Heterogeneity was assessed via Q-test and I² statistic: a fixed-effects model was used if P > 0.1 and I²<50% (low heterogeneity); otherwise, a random-effects model was applied. Sensitivity analysis was performed for predictors with high heterogeneity. Odds ratios (OR) with 95% confidence intervals (CI) were used as effect measures.

Results

Literature screening outcomes

After rigorous screening, 14 studies were included: 12 in Chinese, 2 in English, and 13 published within the past 5 years. All studies adopted a retrospective design, with HF incidence ranging from 3.0% to 37.5%. Baseline characteristics of included studies are summarized in Table 1.

Table 1. Basic characteristics of included studiesAuthor (year)Country/RegionStudy TypeStudy DesignStudy SubjectsData SourceHF Incidence (%) Zhang 2024^[13]^Hunan, ChinaD + VretrospectiveFirst PCI, age > 18Hospital EHR system + China Chest Pain Center Database3.0 Song 2024^[14]^Guangxi, ChinaD + VretrospectivePCI, age ≥ 18Hospital EHR systemA:12.2, B:16.2 Guo 2024^[15]^Beijing, ChinaD + VretrospectivePCI, age ≥ 60Hospital EHR systemA:33.3, B:37.5 Qian 2024^[16]^Beijing, ChinaD + VretrospectivePCI, age ≥ 60Hospital EHR systemA:31.5, B:32.1 Chen 2024^[17]^Zhejiang, ChinaDretrospectivePost-PCI patientsHospital EHR system16.1 Yang 2024^[18]^Nanchong, ChinaD + VretrospectiveFirst PCI, age ≥ 18Clinical medical records28.5 Chen 2024^[19]^Wuhan, ChinaD + VretrospectivePCI, age > 18Hospital EHR systemA:5.4, B:8.5 Zhou 2022^[20]^Anhui, ChinaD + VretrospectivePCI, age ≥ 18Hospital EHR system29.81 Tang 2022^[21]^Guangdong, ChinaDretrospectivePCI, age > 18Hospital EHR system31.25 Wang 2022^[22]^Zhejiang, ChinaDretrospectivePCI, age ≥ 18Hospital EHR system34.67 Fu 2021^[23]^Yunnan, ChinaD + VretrospectivePCI, age ≥ 18Hospital EHR system25.8 Li 2021^[24]^Guangdong, ChinaD + VretrospectivePost-PCI patientsHospital EHR system34.16 Sun 2020^[25]^Gansu, ChinaD + VretrospectivePost-PCI patientsHospital EHR system28.0 Li 2019^[26]^Tianjin, ChinaD + VretrospectivePost-PCI patientsHospital EHR system30.1A: Modeling cohort; B: Validation cohort; D: Model development-only; D + V: Model development and validationD + V: Model development and validation, EHR Electronic health record

Model construction details

All 14 studies developed risk prediction models, with total sample sizes ranging from 139 to 715 participants. Key parameters (e.g., variable selection, modeling methods, and sample distribution) are presented in Table 2.

Table 2. Construction and predictive performance of HF risk prediction models for AMI patients Post-PCILiteratureVariable Selection MethodModeling MethodSample SizeModel PerformanceValidation MethodMissing Data HandlingContinuous Variable HandlingModeling/ValidationAUC/CICalibration Zhang 2024^[13]^Univariate analysisLogistic regression664/-0.847/-H-L test,P = 0.938Bootstrap internal-b Song 2024^[14]^Univariate analysisLasso regression, logistic regression473/2030.87/-H-L test,P = 0.3036Internal-b Guo 2024^[15]^Univariate analysisLogistic regression222/560.897/0.567H-L test,P = 0.545Bootstrap internal-a Qian 2024^[16]^Univariate analysisLogistic regression111/280.903/0.573H-L test, P = 0.532Bootstrap internal-a Chen 2024^[17]^Univariate analysisLogistic regression193/-0.893/0.856Calibration curve--a Yang 2024^[18]^Univariate analysisLasso regression, logistic regression288/-0.894/0.862Calibration curveBootstrap internalDelete directlya Chen 2024^[19]^Lasso regressionLogistic regression503/2120.966/-Calibration curveBootstrap internal-b Zhou 2022^[20]^Univariate analysisLogistic regression208/-0.856/-Calibration curveInternal-a Tang 2022^[21]^Univariate analysisLogistic regression160/--/0.769Calibration curve--a Wang 2022^[22]^Univariate analysisLogistic regression323/--/0.821---a Fu 2021^[23]^Stepwise backward methodLogistic regression438/134-/0.832H-L test, P = 0.288External-a Li 2021^[24]^Univariate analysisFull subset regression, random forest445/-0.855/0.846Calibration curveInternal-b Sun 2020^[25]^Lasso-logistic regressionLogistic regression318/-0.911/0.911Calibration curveExternal-a Li 2019^[26]^Stepwise backward methodLogistic regression495/128-/0.81Calibration curveExternal-aAUC Area under the receiver operating characteristic curve,* CI* Concordance index, H-L**test Hosmer-Lemeshow goodness-of-fit testa: Converted to categorical variables; b: Maintained as continuous variables

Model predictive performance

Model performance was evaluated by discriminative ability (AUC/C-index) and calibration (H-L test, calibration curves). All 14 studies reported discriminative ability, with AUC ranging from 0.847 to 0.966 (indicating strong ability to distinguish high/low-risk patients). One study omitted calibration data; the remaining 13 used either H-L tests (P > 0.05, indicating good calibration) or calibration curves (visually consistent with ideal predictions). Data are presented in Table 2.

Model presentation

All 14 models were presented as nomograms—an intuitive format favored for clinical use. Key predictors, applicability, and limitations of each model are summarized in Table 3.

Table 3. Overview of HF risk prediction models for AMI patients Post-PCILiteratureNo. of PredictorsModel FormatKey PredictorsApplicability & Limitations Zhang 2024^[13]^3NomogramKillip classification, renal insufficiency, troponin TGood applicability; missing potential influential factors. Song 2024^[14]^5NomogramAge, troponin, D-dimer, LVEF, Gensini scoreGood applicability; lacks external validation. Guo 2024^[15]^4NomogramNT-proBNP, NLR, NHR, MSIGood applicability; small sample size (needs large-scale validation). Qian 2024^[16]^4NomogramNT-proBNP, NLR, NHR, MSIGood applicability; small sample size, limited predictors. Chen 2024^[17]^4NomogramTroponin I, left ventricular global longitudinal strain, global wasted work, global work indexGood applicability; single-center design (needs external validation). Yang 2024^[18]^7NomogramAnterior descending artery lesion, Gensini score, troponin T, serum creatinine, hs-CRP, total bilirubin, ventricular wall motion amplitudeGood applicability; lacks external validation. Chen 2024^[19]^6NomogramAge, hypertension prevalence, CCU admission rate, hospital stay duration, onset-to-admission time, D-dimerGood applicability; lacks multicenter validation. Zhou 2022^[20]^13NomogramAge, diseased vessels, anterior MI, Gensini score, hypertension, diabetes, LVEF, troponin I, CK-MB, NT-proBNP, hs-CRP, revascularization, ventricular wall motion amplitudeGood applicability; single-center, small sample size, long data collection period. Tang 2022^[21]^6NomogramAge, Killip classification, blood urea nitrogen, triglycerides, LDL-C, HDL-CGood applicability; single-center, small sample size. Wang 2022^[22]^7NomogramAge, diabetes, cardio-cerebrovascular disease, Gensini grade, hs-CRP, serum creatinine, troponin TGood applicability; lacks multicenter validation. Fu 2021^[23]^7NomogramAge, Gensini score grade, arrhythmia, cardiac structural changes, serum creatinine, troponin T, ventricular wall motion amplitudeGood applicability; single-center (needs prospective validation). Li 2021^[24]^6NomogramLVEF, age, white blood cell count, glomerular filtration rate, plasma fibrinogen, heart rateGood applicability; unproven long-term predictive ability. Sun 2020^[25]^8Nomogram + Simplified ScoreGender, age, 24 h pulmonary rales, onset-to-treatment time, BNP, LVEF, atrial fibrillation, diseased vesselsEasy to use; small sample size (needs further validation for promotion). Li 2019^[26]^7NomogramAge, Gensini score, post-infarction arrhythmia, troponin T, serum creatinine, cardiac structural changes, ventricular wall motion amplitudeGood applicability; needs prospective validation.LVEF Left ventricular ejection fraction, NT-proBNP N-terminal pro-brain natriuretic peptide, NLR Neutrophil-to-lymphocyte ratio, NHR Neutrophil-to-high-density lipoprotein cholesterol ratio, MSI Modified shock index, hs-CRP High-sensitivity C-reactive protein, CK-MB Creatine kinase-MB, LDL-C Low-density lipoprotein cholesterol, HDL-C High-density lipoprotein cholesterol, CCU Coronary care unit

Bias risk and practicality assessment

The PROBAST tool was used to evaluate the risk of bias and applicability of the 14 included studies, and the evaluation results were carefully reviewed to ensure accuracy. In the study population domain, all studies were rated as high risk of bias due to their retrospective design; in the predictor and outcome domains, all were rated as low risk of bias; in the statistical analysis domain, all were rated as high risk of bias. In the applicability evaluation domain, all 14 studies were rated as low risk in all domains and overall evaluation. See Table 4 for details.

To further evaluate the clinical practical value of the included models, this study developed a practicality grading standard based on four core dimensions: completeness of external validation, adequacy of sample size/number of events, accessibility of predictive variables, and convenience of deployment, and graded the 14 models. The specific characteristics and grading results are shown in Table 5.

Table 4. Risk of bias and applicability evaluation of included studiesLiteratureRisk of BiasApplicabilityOverallStudy PopulationPredictorsOutcomeAnalysisStudy PopulationPredictorsOutcomeBiasApplicability Zhang 2024^[13]^+--++++-+ Song 2024^[14]^+--++++-+ Guo 2024^[15]^+--++++-+ Qian 2024^[16]^+--++++-+ Chen 2024^[17]^+--++++-+ Yang 2024^[18]^+--++++-+ Chen 2024^[19]^+--++++-+ Zhou 2022^[20]^+--++++-+ Tang 2022^[21]^+--++++-+ Wang 2022^[22]^+--++++-+ Fu 2021^[23]^+--++++-+ Li 2021^[24]^+--++++-+ Sun 2020^[25]^+--++++-+ Li 2019^[26]^+--++++-+Note: +: Low risk of bias, high applicability; -: High risk of bias, low applicability

Table 5. Grading evaluation table of model practicalityModel(Author/Year)External ValidationSample Size/Number of Events(Modeling Cohort)Accessibility of Predictive Variables(Routine Test = 1, Special Test = 0)Deployment MethodPracticality Grade Zhang 2024^[13]^None664/203/3 (Killip Classification, Renal Insufficiency, Troponin T)NomogramMedium Song 2024^[14]^None473/585/5 (Age, Troponin, D-dimer, LVEF, Gensini Score)NomogramMedium Guo 2024^[15]^None222/744/4 (NT-proBNP, NLR, NHR, MSI)NomogramLow Qian 2024^[16]^None111/354/4 (NT-proBNP, NLR, NHR, MSI)NomogramLow Chen 2024^[17]^None193/313/4 (Troponin I, GLS requiring special ultrasonic analysis)NomogramLow Yang 2024^[18]^None288/826/7 (Total Bilirubin requiring biochemical testing)NomogramMedium Chen 2024^[19]^None503/456/6 (CCU Admission History, Length of Hospital Stay)NomogramMedium Zhou 2022^[20]^None208/6212/13 (Revascularization History)NomogramMedium Tang 2022^[21]^None160/506/6 (Triglyceride, LDL-C)NomogramLow Wang 2022^[22]^None323/1127/7 (Cardio-Cerebrovascular Disease History, Gensini Classification)NomogramMedium Fu 2021^[23]^Yes438/1137/7 (Arrhythmia, Cardiac Structural Changes)NomogramHigh Li 2021^[24]^None445/1526/6 (White Blood Cell Count, Glomerular Filtration Rate, Fibrinogen)NomogramMedium Sun 2020^[25]^Yes318/898/8 (24-hour Pulmonary Rales on Admission, Atrial Fibrillation History)Nomogram + Simplified Scoring TableHigh Li 2019^[26]^Yes495/1497/7 (Post-infarction Arrhythmia, Cardiac Structural Changes)NomogramHighGrading Criteria:High: Meets the criteria of "having external validation + sample size >400/number of events >100 + all predictive variables are routine tests + convenient deployment method"Medium: Meets the criteria of "no external validation but with internal validation + sample size 200–400/number of events 50–100 + no less than 80% of the predictive variables are routine tests + nomogram only"Low: Meets the criteria of "no validation + sample size < 200/number of events < 50 + no less than 20% of the predictive variables are special tests + nomogram only"

Meta-analysis of predictors

Further meta-analysis of common predictors in each model showed that Killip classification, age, Gensini score, D-dimer level, NT-proBNP, NHR, NLR, MSI, troponin I, cardiac troponin T, serum creatinine, hs-CRP, ventricular wall motion amplitude, hypertension, time from onset to treatment, number of diseased vessels, diabetes mellitus, arrhythmia, and cardiac structural changes were effective predictors of HF after PCI in AMI patients (P < 0.05). Sensitivity analysis was performed for predictors with high heterogeneity, and the combined statistical results showed no significant changes, indicating that the meta-analysis results were relatively stable; see Table 6 for details.

Table 6. Meta-Analysis results of HF predictors in AMI patients Post-PCIPredictorNo. of StudiesHeterogeneityEffect ModelMeta AnalysisI^2^(%)P ValueOR(95%CI)P ValueKillip classification200.95fixed3.06(1.68,5.56)< 0.001Age992< 0.001random1.77(1.29,2.41)< 0.001Gensini score691< 0.001random7.06(1.73,28.75)0.006LVEF490< 0.001random1.02(0.75,1.39)0.910D-dimer level200.70fixed2.64(1.56,4.47)< 0.001NT-proBNP300.89fixed10.49(5.26,20.91)< 0.001NHR200.94fixed9.25(3.05,28.01)< 0.001NLR201.00fixed10.83(3.51,33.43)< 0.001MSI200.99fixed8.27(2.67,25.57)< 0.001Troponin I293< 0.001fixed1.36(1.15,1.60)< 0.001Cardiac troponin T4710.01random1.42(1.01,2.00)0.04Serum creatinine4390.18random6.67(2.42,18.42)< 0.001hs-CRP3500.13random10.22(2.32,45.00)0.002Ventricular wall motion amplitude4510.10random5.37(1.90,15.15)0.002Hypertension2380.20fixed4.27(2.11,8.65)< 0.001Onset-to-treatment time2510.15fixed2.24(1.43,3.52)< 0.001Diseased vessels2700.07fixed5.46(2.63,11.36)< 0.001Diabetes mellitus200.51fixed4.50(2.28,8.91)< 0.001Arrhythmia200.81fixed2.05(1.33,3.16)0.001Cardiac structural changes200.95fixed2.43(1.39,4.23)0.002LVEF Left ventricular ejection fraction, NT-proBNP N-terminal pro-brain natriuretic peptide, NLR Neutrophil-to-lymphocyte ratio, NHR Neutrophil-to-high-density lipoprotein cholesterol ratio, MSI Modified shock index, hs-CRP High-sensitivity C-reactive protein

Discussion

Comprehensive evaluation of model predictive performance, risk of bias, and clinical practicality

This study systematically reviewed the latest research on risk prediction models for HF after PCI in AMI patients. The results showed that existing models exhibit promising predictive potential in terms of discriminative ability (AUC: 0.847~0.966), indicating their good ability to identify high-risk patients. However, the evaluation of model quality and transformation value needs to go beyond simple discriminative indicators. The PROBAST tool assessment revealed a key contradiction: despite high predictive performance, all included studies had high risk of bias in the "study population" (all retrospective designs) and "statistical analysis" (such as insufficient sample size, improper handling of continuous variables, and single variable selection method), which seriously restricts the reliability and extrapolatability of model results.

Further comprehensive analysis combined with model practicality grading (Table 5) clearly outlined the three-tier structure of the current research ecosystem:

Top tier (High practicality models, 3 studies): Common features include completion of external validation, relatively sufficient sample size/number of events, all predictive variables being routine clinical tests, and convenient deployment. These set a preliminary benchmark for "usable" and even "easy-to-use" models in the future.

Middle tier (Medium practicality models, 7 studies): Constituting the main body of the research, most conducted internal validation and adopted nomogram presentation. However, due to the lack of external validation, limited sample size, or poor accessibility of some variables, their clinical promotion value is uncertain.

Bottom tier (Low practicality models, 4 studies): Limited by small sample size, lack of validation, and poor accessibility of variables, they are currently mainly confined to methodological exploration.

In summary, this study identified the core challenge in the current field as the"imbalance among predictive performance, risk of bias, and clinical practicality". The grading results in Table 5 clearly show that improving predictive performance alone cannot naturally realize clinical transformation of models. High risk of bias fundamentally undermines the scientific foundation of prediction models, while low practicality directly hinders their integration into clinical workflows.

Therefore, future studies must establish a trinity model construction goal of "high predictive performance, low risk of bias, and high clinical practicality". To achieve this goal, this study proposes a two-dimensional optimization path: first, in terms of study design and methodology, strictly follow the PROBAST and TRIPOD guidelines, and strive to reduce the risk of bias from"high" to "low". Specific measures include prioritizing prospective designs, standardizing the processing of continuous variables and variable selection processes, properly handling missing data, and expanding sample size. Second, in terms of model transformation, actively carry out multicenter external validation, ensure the routine accessibility of predictors, and develop more convenient deployment tools (such as online calculators integrated into electronic medical record systems). Through such systematic efforts that balance methodological rigor and clinical transformation needs, we can promote HF prediction models after AMI-PCI from"high performance on paper" to "high value at the bedside", and ultimately provide effective decision support for improving patient prognosis.

Effective predictors of heart failure after PCI in patients with acute myocardial infarction

Through meta-analysis of common predictors in the included models, this study identified multiple predictors significantly associated with the occurrence of HF after PCI in AMI patients. These predictors include Killip classification, age, Gensini score, D-dimer level, NT-proBNP, NHR, NLR, MSI, troponin I, cardiac troponin T, serum creatinine, hs-CRP, ventricular wall motion amplitude, hypertension, time from onset to treatment, number of diseased vessels, diabetes mellitus, arrhythmia, and cardiac structural changes.

Killip classification and Gensini score: important clinical assessment indicators

As traditional clinical assessment tools, Killip classification and Gensini score are widely used to evaluate the severity of illness in AMI patients. Killip classification is based on patients’ clinical symptoms, and studies have shown [6, 7] that it plays an important role in predicting postoperative HF. A higher Killip classification usually reflects more severe cardiac insufficiency, so Killip classification can effectively predict the risk of HF. In addition, Gensini score indirectly reflects the degree of myocardial injury by assessing the severity of coronary artery lesions. A higher Gensini score indicates more severe coronary artery lesions in patients, which usually leads to greater myocardial damage and a higher risk of HF [27–30]. Therefore, Gensini score is also of great significance in predicting HF after PCI.

Diagnostic value of biomarkers

In recent years, biomarkers have gradually become important tools for predicting HF in AMI patients. NT-proBNP, as a sensitive marker of HF, has been extensively studied and proven to have high sensitivity and specificity in predicting HF. The level of NT-proBNP is closely related to the severity of congestive heart failure, and an increase in NT-proBNP after surgery usually indicates the occurrence of HF [31–34]. Therefore, NT-proBNP monitoring after PCI has important clinical significance. D-dimer and hs-CRP are biomarkers reflecting inflammatory response and coagulation status, and their increased levels are also closely related to the occurrence of HF. An increase in D-dimer suggests that the patient may have thrombosis, which further aggravates HF symptoms [35], while hs-CRP reflects the chronic inflammatory response in the body, and these factors may increase the risk of HF [36].

Impact of clinical characteristics on heart failure

Patients’ underlying diseases, such as age, hypertension, and diabetes mellitus, play an important role in the occurrence of postoperative HF. With the increase of age, the function of the cardiovascular system gradually declines, and the myocardial repair ability weakens, leading to a higher risk of HF in elderly patients [37]. Hypertensive patients are often accompanied by left ventricular hypertrophy, myocardial injury, and changes in cardiac structure. These pathological changes make hypertensive patients more prone to HF after PCI [38]. Diabetes mellitus further increases the burden on the heart through mechanisms such as glucose metabolism disorders and atherosclerosis, leading to an increased risk of HF in patients [39]. Therefore, these patients have a higher cardiovascular risk, and special attention should be paid to the prediction and management of postoperative HF.

Role of inflammatory response

NHR (neutrophil-to-high-density lipoprotein cholesterol ratio) and NLR (neutrophil-to-lymphocyte ratio) are inflammatory response markers that have attracted attention in recent years. Studies have shown that their increase is closely related to the risk of HF after acute myocardial infarction [40–43]. An increase in neutrophils usually indicates a strong inflammatory response in the body, which can aggravate myocardial injury and further induce HF [44]. An increase in NHR and NLR may indicate a strong inflammatory response in patients, increasing the risk of postoperative HF. Therefore, these indicators have certain early warning value as inflammatory assessment tools in clinical practice.

Changes in cardiac function and structure

Troponin I and troponin T are common markers of myocardial injury. An increase in these markers after surgery usually indicates further myocardial injury. The increase of troponin can reflect the existence of myocardial injury in the early stage of acute myocardial infarction and is closely related to the occurrence of HF [45, 46]. In addition, an increase in serum creatinine often reflects renal insufficiency. There is an interaction between renal and cardiac functions, and renal failure may aggravate HF symptoms. Therefore, monitoring serum creatinine levels is of great significance for predicting the risk of HF after PCI [47, 48]. Changes in ventricular wall motion amplitude can also be used as an indicator of cardiac function injury. A decrease in ventricular wall motion amplitude shown by cardiac ultrasound often means impaired myocardial systolic function, which may indicate the occurrence of HF [49–51].

Among them, Killip classification and Gensini score are important indicators reflecting patients’ cardiac function and the severity of coronary artery lesions, which are closely related to the occurrence of HF. Patient-related factors such as age, hypertension, and diabetes mellitus are also important predictors. These factors may increase the risk of HF by affecting vascular function and myocardial repair ability. In addition, biomarkers such as NT-proBNP, D-dimer, and hs-CRP show high sensitivity and specificity in predicting HF, suggesting that these indicators have important application value in clinical practice.

Future research directions and implications

Although existing models show good performance in predicting the risk of HF after PCI in AMI patients, they still have limitations and need to be optimized in the following directions.

Research progress and application of artificial intelligence in China

Some studies have used machine learning algorithms such as Lasso regression and random forest to construct models, and their predictive performance is better than traditional logistic regression. The latest research published by Chinese scholars in 2025 further focuses on model interpretability, coverage of subpopulations, and clinical transformation and implementation, providing more solid local evidence for the application of AI technology in predicting HF after AMI-PCI:

A single-center retrospective study by Guo et al. (n = 1574 AMI patients) first applied the TabNet deep learning model to predict the severity of HF (Killip classification) after surgery, and compared the performance of random forest (RF), XGBoost, and multi-layer perceptron (MLP) models. The results showed that the TabNet model performed the best: the AUC for Killip four-classification prediction was 0.827, and for two-classification prediction was 0.831, which was significantly better than traditional machine learning models. The study identified GRACE score, TIMI score, NT-proBNP, LVEF, and creatinine clearance rate (CCR) as key predictors through the SHAP method. Among them, increased NT-proBNP and decreased LVEF were significantly associated with higher Killip classification, which was completely consistent with the conclusion of this meta-analysis that “NT-proBNP and LVEF are effective predictors”; at the same time, a directly clinically applicable web platform was developed and made public (https://prediction-killip-gby.streamlit.app/), realizing the closed loop of “model construction - interpretation - clinical transformation” [52].

A single-center prospective study by Wang et al. focused on the special population of premature myocardial infarction (PMI, males ≤ 50 years old, females ≤ 55 years old) and constructed a prediction model for in-hospital heart failure with preserved ejection fraction (HFpEF). The study included 840 PMI patients (268 developed HFpEF, with an incidence of 31.90), and compared 5 algorithms including Lasso-Logistic, XGBoost, and random forest. The results showed that the XGBoost model performed the best, significantly better than other models. The model finally included 10 key predictors, including BNP >100 pg/ml, SYNTAX score >14.5, age, and monocyte-to-lymphocyte ratio (MLR) >0.3. Among them, increased BNP and complex coronary lesions were strongly associated with an increased risk of HFpEF, which was highly consistent with the conclusion of this meta-analysis that “biomarkers and the degree of coronary artery lesions are core predictors”; at the same time, a visual prediction system was developed (https://hfpefpmi.shinyapps.io/apppredict/), which supports clinicians to quickly obtain the probability of HFpEF and SHAP explanation diagrams after inputting patients’ clinical indicators, filling the gap in predicting specific HF subtypes in young AMI patients [53].

A single-center retrospective study by Lin et al. constructed a prediction model for the 3-year risk of HF after PCI in AMI patients, including 1220 patients (244 developed HF, with an incidence of 20%). Six key factors (LDH, CK-MB, hsCRP, NT-proBNP, LVEF, and left ventricular end-systolic diameter (LVDs)) were selected through LASSO regression, and the performance of XGBoost, random forest, SVM, and logistic regression models was compared. The results showed that the XGBoost model performed the best. SHAP analysis identified LVEF, LVDs, and LDH as the three most important predictors. Among them, the risk of HF was significantly increased when LVEF < 50%, LVDs >4.0 cm, and LDH >650 U/L, which was consistent with the conclusion of this meta-analysis that “LVEF, myocardial injury markers, and cardiac structural changes are effective predictors”; at the same time, individual-level prediction interpretation was realized through SHAP waterfall diagrams and force-directed diagrams, clearly showing the contribution of each factor to the patient’s risk, and providing an intuitive tool for clinical personalized assessment [54].

The above studies in 2025 show three core advantages: first, stronger algorithm adaptability, including deep structured models such as TabNet (suitable for multi-classification tasks) and mature integrated algorithms such as XGBoost (suitable for subtype-specific and long-term risk prediction), which can meet the needs of different clinical scenarios; second, comprehensive enhancement of model interpretability, all using the SHAP method to reveal the prediction logic, and the core predictors are highly consistent with the results of this meta-analysis, further verifying clinical generality; third, prominent clinical transformation and implementation, two studies developed publicly accessible visual platforms, which can quickly obtain prediction results without complex operations, reducing the threshold for clinical application.

At the same time, limitations should be viewed objectively: all three studies were single-center designs, without multicenter external validation, and the extrapolatability of the models in different medical environments needs to be verified; the predictors are still mainly structured clinical indicators and biomarkers, without fully integrating unstructured data such as radiomics and genomic data, which may miss potential prediction information; there is no targeted prediction tool for complex populations with multiple underlying diseases. Future studies can promote multicenter prospective validation of models on the existing basis, explore the integration mode of “structured data + multi-modal data”, and develop special models focusing on more complex subpopulations to further improve the generality, accuracy, and clinical transformation value of AI models.

Optimization of study design and methodology

All 14 existing studies were conducted by Chinese scholars, and the model presentation method was nomogram. Although nomogram is an intuitive risk prediction tool easily accepted by clinicians, its generality and extrapolatability still need to be verified. Current studies are mostly single-center retrospective designs, lacking external validation across regions and ethnic groups, which limits the clinical application scope of the models. In the future, it is necessary to strictly follow the PROBAST tool and TRIPOD statement to optimize study design: prioritize multicenter, prospective cohort studies to expand sample size and cover different regional medical environments; standardize the processing of continuous variables (verify the linear relationship before deciding whether to classify), missing data (adopt multiple imputation instead of direct deletion), and variable selection (avoid relying solely on univariate analysis); explore the differences in HF risk among different regions and medical resource conditions for the Chinese population, and construct more generalizable prediction models.

Model presentation and clinical transformation

Existing models are mainly nomograms. In the future, the integration mode of “AI algorithm + nomogram” can be explored to retain the intuitiveness of nomograms while combining the advantages of AI in processing high-dimensional data to improve prediction accuracy; in addition, models can be combined with dynamic risk scoring systems and hospital electronic medical record systems to develop convenient clinical decision support tools, enhancing the practicality and user experience of models; at the same time, international cooperation should be strengthened to include patient data from different ethnic groups and regions, verify the applicability of models in global populations, and promote the international application of models.

Limitations of the study

This study has the following limitations, which should be carefully considered when interpreting the results and conclusions:

Limitations in literature search and representativeness of included populations: This study only searched Chinese and English databases and published literature, not including gray literature and studies in other languages, which may lead to potential publication bias and literature omission. In addition, all 14 included studies were from China, lacking verification in populations of different ethnic groups, regions, and medical systems. Therefore, the predictors integrated by the meta-analysis of this study and the model practicality evaluation results have limited extrapolatability worldwide.

Insufficient breadth and depth of model validation: Among the included models, only 3 (21.4%) were rated as “high practicality” and completed external validation, and most of these validations were conducted in single centers or small scopes, lacking extensive, prospective multicenter validation across different geographical regions and medical institutions at all levels. The remaining 11 models (78.6%) did not undergo external validation, and their predictive performance and stability are highly dependent on internal validation results, making the existing evidence system’s support for the performance of models in real-world clinical environments still weak.

Methodological quality and risk of bias: According to the PROBAST tool assessment, all included studies had high risk of bias, mainly due to the inherent limitations of retrospective study design and common methodological deficiencies in some studies, such as insufficient sample size and number of outcome events (failing to meet the empirical criterion of “at least 10–20 events per predictive variable” in prediction model development), over-reliance on univariate analysis for initial variable screening, artificial classification of continuous variables without pre-testing their linear relationship with outcomes, and unclear or improper handling of missing data. These factors may affect the accuracy and reliability of model parameter estimation.

Imperfection of practicality evaluation criteria: This study independently constructed a model practicality grading standard based on four dimensions: external validation, sample size, variable accessibility, and deployment method. Although the standard strives to be objective, it has not been cross-validated with existing international model practicality or implementability assessment tools, and its assessment validity needs further confirmation in future studies. At the same time, the current assessment mainly focuses on the “theoretical practicality” at the model construction and reporting level, and has not yet evaluated the “real-world impact” of models on diagnosis and treatment decisions and patient prognosis after clinical application. The evaluation dimensions need to be expanded in the future.

Limitations of predictors and model functions: The effective predictors identified through meta-analysis are mainly traditional clinical characteristics, routine laboratory biomarkers, and basic imaging parameters, lacking exploration and integration of new multi-omics indicators such as radiomics and genetic markers. In addition, all included models aim to predict the binary outcome of “whether HF occurs”, and cannot perform more refined risk stratification prediction on the severity, subtype, or occurrence time of HF, thus limiting their potential value in individualized precise management to a certain extent.

Timeliness gap between research data and clinical practice: The original medical record data of some included studies can be traced back to earlier years (as early as before 2019). In recent years, the acute phase management strategies, PCI technology, and guideline-directed pharmacotherapy for acute myocardial infarction have been continuously optimized and updated. These changes in diagnosis and treatment models may potentially alter the incidence of HF and the strength of its association with traditional predictors. Therefore, the applicability and predictive performance of some models constructed based on historical data may need to be re-evaluated and updated in the context of current best clinical practice.

Conclusion

Through systematic review and meta-analysis, this study comprehensively evaluated the existing risk prediction models for HF after PCI in AMI patients. The results showed that existing models exhibit relatively ideal predictive potential in terms of discriminative ability, and meta-analysis also identified multiple predictors significantly associated with the occurrence of HF, providing valuable references for clinical identification of high-risk populations. However, the PROBAST tool assessment indicated that all included studies had a high risk of bias, mainly due to methodological limitations such as retrospective design, insufficient sample size, irregular variable handling, and inadequate validation. In addition, model practicality grading showed that only a few models have good clinical transformation potential.

In summary, although existing models show certain predictive performance, their methodological quality and clinical extrapolatability still need to be improved. Future studies should strictly follow the relevant specifications for prediction models, and strive to systematically reduce the risk of bias through prospective design, multicenter cooperation, standardized statistical methods, expanded sample size and number of events, and strengthened internal and external validation. At the same time, attention should be paid to the clinical accessibility of predictors and the convenience of model deployment, promoting the transformation of models from methodological construction to real-world application, so as to ultimately provide clinical practice with scientific, robust, and practical risk assessment tools to assist in the prevention, treatment, and individualized management of HF after PCI in AMI patients.

Supplementary Information

Supplementary Material 1.

Supplementary Material 2.

Supplementary Material 3.

Supplementary Material 4.

Supplementary Material 5.