Development and real-world cohort validation of a meta-analysis-derived simplified scoring model for cardiac resynchronisation therapy response: a study protocol

Abstract

Background

Cardiac resynchronisation therapy (CRT) is a cornerstone device-based treatment for patients with heart failure with ventricular dyssynchrony. However, approximately 30–40% of recipients fail to achieve clinical response. Despite extensive research, validated prediction tools grounded in high-level evidence and readily applicable in clinical practice remain lacking. This study protocol describes the development and real-world validation of a simplified clinical scoring model for CRT response derived from systematic review and meta-analysis.

Methods and analysis

This study will develop a CRT response prediction model via meta-analysis and preliminarily validates it in a single-centre retrospective cohort. Initially, systematic searches of multiple databases up to 31 January 2026 and meta-analysis will synthesise effect estimates for candidate predictors, creating an evidence-based foundation that conceptually functions as a ‘training dataset’. Predictor selection and prioritisation will be guided by study frequency, effect magnitude and clinical accessibility, with factor weights derived directly from pooled random-effects meta-analytic estimates. Log relative risks will be converted to integer scores to establish a series of nested prediction models. Model performance will then be comprehensively assessed in an independent ‘validation dataset’ comprising a single-centre cohort from Xinjiang Medical University, evaluating discrimination (area under the receiver operating characteristic curve), calibration (calibration plots and Hosmer-Lemeshow test) and clinical utility (decision curve analysis). The final scoring system will be identified through comparative model evaluation guided by parsimony principles.

Ethics and dissemination

This meta-analysis exclusively uses published, de-identified data and therefore does not require ethical approval. The validation cohort employs retrospectively anonymised patient data in strict adherence to the ethical principles of the Declaration of Helsinki. The study protocol has been approved by the Institutional Review Board of the First Affiliated Hospital of Xinjiang Medical University (Approval No.: K202403-48-2503A-Y1). As this constitutes a retrospective analysis of existing data, individual informed consent will be waived. Comprehensive measures to protect participant privacy and ensure data integrity will be implemented throughout all research procedures. The findings will be presented at academic gatherings or published in scholarly, peer-reviewed journals.

PROSPERO registration number

CRD42024572313

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• Meta-analysis will quantify critical cardiac resynchronisation therapy (CRT) response predictors as pooled effect estimates with 95% CI to establish an evidence-based predictive framework.

• Comparing nested models will optimise performance-parsimony balance, yielding a bedside CRT non-response probability reference table for clinical decision support.

• Despite prioritising multivariable-adjusted estimates, model weights derived from observational meta-analyses remain subject to residual confounding. Furthermore, this approach inherently cannot address predictor interactions or multicollinearity; although we empirically examine incremental predictor value through nested model comparisons in the independent validation cohort, these limitations cannot be fully eliminated.

• Single-centre retrospective data provide only preliminary validation evidence, which inherently constrains the model’s external validity and widespread clinical applicability. Multicentre prospective validation is essential for clinical implementation.

Introduction

Chronic heart failure (CHF) represents the terminal stage of progressive cardiac dysfunction, affecting 1–3% of the adult population. Heart failure with reduced ejection fraction accounts for approximately half of these cases.¹ While substantial advances in pharmacotherapy have occurred, with guideline-directed medical therapy (GDMT) significantly improving outcomes, optimisation of drug regimens remains challenging due to multiple patient-specific factors.² Consequently, a considerable proportion of patients fail to achieve optimal response despite maximal GDMT.² For end-stage CHF, advanced options including left ventricular assist device³ implantation and cardiac transplantation⁴ are available, but their clinical application is limited by donor scarcity, technical complexity and substantial healthcare costs.

The natural history of CHF involves progressive ventricular remodelling, cardiac enlargement and resultant electromechanical dyssynchrony.⁵ Physiological pacing strategies—including cardiac resynchronisation therapy (CRT) and conductive system pacing (CSP) modalities such as His-bundle pacing and left bundle branch area pacing—play a pivotal role in improving or maintaining ventricular synchrony, establishing them as effective therapeutic interventions for CHF.⁵ However, despite recent heightened interest in CSP, it has not yet supplanted biventricular pacing (BiVP) as the standard of care, pending validation through large-scale prospective trials. Accordingly, BiVP remains endorsed as first-line therapy by major international guidelines.⁵

Cumulative evidence demonstrates that 30–40% of patients experience suboptimal CRT response, defined by lack of improvement in clinical outcomes and symptom burden.⁶ Given CRT’s substantial costs and procedural risks, pre-implantation identification of likely responders constitutes a critical clinical priority for both individualised therapy and resource optimisation. However, response prediction demands multivariable integration rather than reliance on isolated parameters. Prior investigations have examined diverse candidate predictors spanning demographics, comorbidities, electrocardiographic, echocardiographic and cardiac magnetic resonance parameters, yet most lack validation in large-scale cohorts.⁷ Although one systematic review compiled prognostic factors, it omitted meta-analytic synthesis, precluding rigorous quantitative evidence evaluation.⁸ Existing prediction models demonstrate substantial heterogeneity in design, outcome definitions, feature dimensions and algorithmic approaches, limiting their generalisability.⁹ Furthermore, models derived from small samples or complex artificial intelligence algorithms, while achieving high internal validation performance, face significant challenges regarding clinical utility in broader populations.⁷ Accordingly, this study protocol aims to identify robust CRT response predictors through meta-analysis and construct and validate an evidence-based, clinically accessible scoring model to support rational decision-making and precision medicine implementation.

Methods and analysis

Meta-analysis

Data sources and search strategy

The meta-analysis protocol was prospectively registered with PROSPERO (CRD42024572313; 31 July 2024). Study findings will be reported following Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines.¹⁰ Candidate predictors of CRT response will be identified through comprehensive systematic review of published studies. Effect estimates—including odds ratios (OR), relative risks (RR) and 95% CIs—will be synthesised via random-effects meta-analysis.

Comprehensive electronic searches will be conducted from database inception to 31 January 2026, across English-language databases (PubMed, EMBASE, Cochrane Library, Web of Science) using Medical Subject Headings and keyword combinations. Chinese databases will include Wanfang Data, China Biomedical Literature Database, China National Knowledge Infrastructure and VIP Database. If any potentially important non-English or non-Chinese articles are identified during screening, they will be documented and considered for inclusion in a sensitivity analysis to assess their impact on the pooled estimates. The search strategy is detailed in online supplemental table 1.

Eligibility criteria

Inclusion criteria: (1) population: patients meeting established criteria for CRT implantation; (2) intervention: CRT device implantation (CRT pacemaker (CRTP) or CRT defibrillator (CRTD)) or upgrades from conventional pacemakers/implantable cardioverter defibrillator (ICD) to CRT systems; (3) predictors: studies reporting ≥1 baseline predictor of CRT outcomes with effect estimates (OR/RR) and 95% CI, or sufficient data for calculation; (4) outcome: clear definition of CRT response incorporating reverse cardiac remodelling and clinical improvement; (5) study design: original prospective or retrospective cohort studies.

Exclusion criteria: (1) descriptive studies (eg, cross-sectional surveys, case reports, case series, conference abstracts), reviews, animal studies, meta-analyses/systematic reviews, case-control studies and randomised controlled trials; (2) predictor domain: markers with limited clinical generalisability (molecular biomarkers, global longitudinal strain, cardiac magnetic resonance parameters); (3) data availability: studies without extractable effect estimates, duplicate publications or those with incomplete/substantially missing data.

Study selection and data extraction

Two investigators will independently perform literature screening to ensure objectivity and reproducibility. EndNote 9.1 will be used for reference management, with deduplication conducted through automated processing supplemented by manual verification. Initial screening of titles and abstracts will be followed by full-text review against eligibility criteria (figure 1). One investigator will perform data extraction using standardised Microsoft Excel 2021 forms, with verification by a second investigator. Extracted variables will include the following: publication characteristics, study location, design, population source, sample sizes (exposed and unexposed groups), demographic and baseline clinical data, outcome definitions, follow-up duration, predictor specifications, event counts with incidence rates, and effect estimates (OR, RR) with 95% CI. Discrepancies will be resolved through consensus discussion; unresolved disagreements will be adjudicated by a third expert investigator, whose decision is final. If third-party arbitration is required for >5% of screened records, the entire selection process will undergo re-evaluation to ensure protocol adherence. A detailed list of articles that required arbitration or consensus discussion during the full-text review stage, along with the reasons for exclusion (if ultimately excluded), will be provided in the Supplementary Material.

Figure 1. Literature screening process. Adapted from Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ 2021;372:n71. doi: 10.1136/bmj.n71. This diagram illustrates the flow of information through the different phases of this systematic review, detailing the numbers of records identified, screened, assessed for eligibility and included, along with reasons for exclusions. Final numbers will be reported upon review completion.

Methodological quality assessment

Methodological quality of included cohort study will be evaluated using the Newcastle-Ottawa Scale (NOS) (https://www.ohri.ca/programs/clinical_epidemiology/oxford.asp). This validated nine-point instrument assesses three domains: participant selection (four points), between-group comparability (two points) and outcome/exposure ascertainment (three points). Quality classifications will follow evidence-based medicine standards: high (≥7 points), moderate (5–6 points) or low (≤4 points). NOS quality assessment criteria are summarised in online supplemental table 2.

Missing data handling

For studies reporting only event counts or incidence rates, risk ratios (RR) and 95% CI will be calculated using established 2×2 contingency table methods.¹¹

$$\mathrm{I}\mathrm{n}\mathrm{c}\mathrm{i}\mathrm{d}\mathrm{e}\mathrm{n}\mathrm{c}\mathrm{e}\mathrm{}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{}\mathrm{o}\mathrm{f}\mathrm{}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{}\mathrm{e}\mathrm{x}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}\mathrm{}\mathrm{g}\mathrm{r}\mathrm{o}\mathrm{u}\mathrm{p}=\frac{\mathrm{a}}{\mathrm{a}+\mathrm{b}}$$

$$\mathrm{I}\mathrm{n}\mathrm{c}\mathrm{i}\mathrm{d}\mathrm{e}\mathrm{n}\mathrm{c}\mathrm{e}\mathrm{}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{}\mathrm{i}\mathrm{n}\mathrm{}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{}\mathrm{n}\mathrm{o}\mathrm{n}-\mathrm{e}\mathrm{x}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}\mathrm{}\mathrm{g}\mathrm{r}\mathrm{o}\mathrm{u}\mathrm{p}=\frac{\mathrm{c}}{\mathrm{c}+\mathrm{d}}$$

$$RR=\frac{\mathrm{I}\mathrm{n}\mathrm{c}\mathrm{i}\mathrm{d}\mathrm{e}\mathrm{n}\mathrm{c}\mathrm{e}\mathrm{}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{}\mathrm{o}\mathrm{f}\mathrm{}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{}\mathrm{e}\mathrm{x}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}\mathrm{}\mathrm{g}\mathrm{r}\mathrm{o}\mathrm{u}\mathrm{p}}{\mathrm{I}\mathrm{n}\mathrm{c}\mathrm{i}\mathrm{d}\mathrm{e}\mathrm{n}\mathrm{c}\mathrm{e}\mathrm{}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{}\mathrm{i}\mathrm{n}\mathrm{}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{}\mathrm{n}\mathrm{o}\mathrm{n}-\mathrm{e}\mathrm{x}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}\mathrm{}\mathrm{g}\mathrm{r}\mathrm{o}\mathrm{u}\mathrm{p}}$$

$$\mathrm{S}\mathrm{E}\left(\mathrm{l}\mathrm{n}\right(RR\left)\right)=\sqrt{\left(\frac{1}{\mathrm{a}}-\frac{1}{\mathrm{a}+\mathrm{b}}\right)+\left(\frac{1}{\mathrm{c}}-\frac{1}{\mathrm{c}+\mathrm{d}}\right)}$$

$$\mathrm{L}\mathrm{o}\mathrm{w}\mathrm{e}\mathrm{r}\mathrm{}\mathrm{l}\mathrm{i}\mathrm{m}\mathrm{i}\mathrm{t}\mathrm{}\mathrm{o}\mathrm{f}\mathrm{}95\mathrm{\%}\mathrm{C}\mathrm{I}=e^{\mathrm{l}\mathrm{n}\left(RR\right)-1.96\times \mathrm{S}\mathrm{E}\left(\mathrm{l}\mathrm{n}\right(\mathrm{R}\mathrm{R}\left)\right)}$$

$$\mathrm{U}\mathrm{p}\mathrm{p}\mathrm{e}\mathrm{r}\mathrm{}\mathrm{l}\mathrm{i}\mathrm{m}\mathrm{i}\mathrm{t}\mathrm{}\mathrm{o}\mathrm{f}\mathrm{}95\mathrm{\%}\mathrm{C}\mathrm{I}=e^{\mathrm{l}\mathrm{n}\left(RR\right)+1.96\times \mathrm{S}\mathrm{E}\left(\mathrm{l}\mathrm{n}\right(\mathrm{R}\mathrm{R}\left)\right)}$$

a: the number of people who fell ill in the exposure group;

b: the number of people who did not fall ill in the exposure group;

c: the number of people who fell ill in the non-exposure group;

d: the number of people who did not fall ill in the non-exposure group.

Pooled effect estimates and heterogeneity assessment

Meta-analysis will be performed for predictors reported with effect sizes extracted hierarchically: (1) multivariable-adjusted estimates, (2) unadjusted multivariable values and (3) univariate estimates. Effect measures will comprise OR or RR with 95% CI. OR values will be converted to approximate RR using the formula RR ≈ OR/[1−P₀+(P₀×OR)]¹² when baseline incidence in the unexposed group (P₀) is available; otherwise, OR will be approximated as RR under the rare outcome assumption. Pooled estimates will be generated using random-effects models (DerSimonian-Laird) with generic inverse-variance weighting. Heterogeneity will be quantified using the I² statistic. Sensitivity analyses will include subgroup analyses by original effect measure type and fixed-effects models (inverse-variance method) to assess robustness.

Publication bias assessment

Publication bias will be evaluated using Egger’s regression test (Stata V.17.0), which conducts weighted linear regression of effect estimates against their precision (SE). A statistically significant intercept (p<0.05) will be interpreted as evidence of funnel plot asymmetry suggestive of publication bias.

Subgroup analysis and exploration of heterogeneity

Subgroup analyses will be conducted for predictors with consistent stratification definitions (identical or similar) across cohorts containing ≥3 studies per stratum. The following prespecified subgroup analyses will be conducted, where data permit: age (≥75 vs <75 years); heart failure aetiology (ischaemic vs non-ischaemic cardiomyopathy); and renal function (estimate the glomerular filtration rate (eGFR) <60 vs ≥60 mL/min/1.73 m²). For each analysis, pooled effect estimates with 95% CI and I² statistics will be calculated.

Where substantial heterogeneity is detected (I² >50% and/or p<0.10 on Cochran’s Q test) and ≥10 studies are available, random-effects meta-regression will be performed to explore potential sources. Candidate moderators will include subgroup analysis variables plus, where data permit: mean age of study population, proportion of male patients, mean baseline left ventricular ejection fraction (LVEF), proportion of patients with left bundle branch block, follow-up duration (months). Each moderator will first be analysed univariately; variables with suggestive associations (p<0.10) will be considered for multivariable models, provided sufficient study numbers exist to avoid overfitting. Meta-regression will use restricted maximum likelihood (REML) estimation, with results presented as regression coefficients, 95% CI and p values. Given the risk of ecological bias, meta-regression findings will be interpreted cautiously as hypothesis-generating rather than confirmatory.

Certainty of evidence assessment

The Grading of Recommendations Assessment, Development and Evaluation (GRADE) framework will be employed to assess the quality of evidence for each pooled predictor outcome. This evaluation will consider five dimensions: risk of bias: assessed using NOS scores of included studies; inconsistency: evaluated through I² statistics and CI overlap; indirectness: judged based on generalisability of study populations, interventions and outcome definitions to the target population; imprecision: assessed through 95% CI width and optimal information size attainment; publication bias: examined using Egger’s regression test and funnel plot visual inspection. For each predictor, overall evidence quality will be rated as high, moderate, low or very low according to standard GRADE guidelines.¹³

Sensitivity analysis

To assess the robustness of meta-analysis findings, a series of sensitivity analyses will be conducted: (1) Leave-one-out analysis: for each predictor meta-analysis, pooled effect estimates will be recalculated sequentially omitting one study at a time to assess disproportionate influence of any single study. (2) Exclusion of high-risk studies: studies with NOS scores ≤4 (indicating high risk of bias) will be excluded to evaluate whether low-quality study inclusion influences pooled estimates. Results will be graphically presented as forest plots in supplementary materials. (3) Fixed-effect models: although random-effects models will serve as the primary approach, fixed-effects models (inverse-variance method) will be employed to evaluate the impact of model selection on pooled estimates and assess conclusion robustness under alternative modelling assumptions. (4) OR-to-RR conversion: studies reporting only OR without sufficient data for RR calculation or conversion will be excluded. Only studies directly reporting RR or providing raw data enabling RR computation will be included. We will compare pooled effect sizes, 95% CI and heterogeneity between RR derived from direct reporting/raw data and RR converted from OR using the aforementioned method. Substantial discrepancies in effect estimates will be discussed regarding their impact on score derivation and interpreted as evidence of potential bias introduced by OR-to-RR conversion. (5) Outcome definition heterogeneity: analysis restricted to studies with CRT response definitions highly consistent with the validation cohort composite endpoint. Pooled effect sizes, 95% CI and heterogeneity from this restricted set will be compared with the primary meta-analysis. Given anticipated limited study numbers meeting these criteria, this analysis will aim to assess the direction and potential magnitude of bias attributable to definition heterogeneity rather than derive alternative effect estimates. Results will be provided in supplementary materials.

Model development

Candidate predictor pool and prioritisation

The initial candidate predictor pool will be derived from meta-analytic results, incorporating robust subgroup and sensitivity analysis findings. Predictors will be prioritised for inclusion according to a hierarchical framework: (1) factors reported in ≥10 independent studies^{14 15} to ensure broad evidence validation; (2) among comparably reported factors, those with larger pooled effect estimates (RR); and (3) guideline-endorsed parameters readily accessible in routine clinical practice (eg, electrocardiographic variables, basic echocardiographic indices, standard laboratory tests).

Nested model construction and score conversion

Based on the prioritisation framework, we will generate a series of nested models (A–K) ranging from simple to complex. Model A includes only the top-ranked predictor; each subsequent model incorporates an additional predictor sequentially (model B adds predictor #2, model C adds predictor #3, etc.).

Model scores will be derived from pooled meta-analytic (RR) using Sullivan’s methodology.¹⁶ The regression coefficient (β) for each predictor equals the natural logarithm of its pooled RR (β=ln[RR]). Each predictor’s clinical score is calculated as 10×β rounded to the nearest integer. For example, a predictor with RR=2.0 (β ≈ 0.693) receives 7 points. Categorical predictors receive the calculated score when present and 0 when absent. Given that the positive outcome is defined as ‘response’, protective factors (RR >1) are assigned positive scores, whereas risk factors (RR <1) receive negative scores.

Model validation

Validation cohort

Data source

The validation cohort will comprise patients who will have undergone CRT implantation (including CRTP, CRTD and ICD-to-CRT upgrades) at the First Affiliated Hospital of Xinjiang Medical University (a third-class first-grade medical centre in China; all clinical workers have obtained professional licences and have undergone specialised training) between January 2018 and October 2025, with ≥6 months of postoperative follow-up.

Sample size calculation

Based on the standard requirement of 15 events per variable for eight candidate predictors,^{17 18} a minimum of ≥120 CRT non-response events is required. Assuming a 35% event rate, the target sample size is calculated as 343 patients.

Eligibility criteria

Inclusion criteria: (1) ≥3 months of optimised GDMT prior to CRT implantation; (2) New York Heart Association (NYHA) functional class II–IV; (3) baseline QRS duration ≥120 ms and LVEF ≤35%; (4) age ≥18 years.

Exclusion criteria: (1) indication for conventional pacemaker or ICD without CRT candidacy; (2) prior CRT device implantation; (3) concurrent severe valvular disease requiring surgical intervention, acute myocardial infarction (<40 days) or non-cardiac conditions with life expectancy <1 year (eg, advanced malignancy); (4) pregnancy or lactation; (5) inability to complete 6-month post-procedural follow-up.

Baseline characteristics

Demographic and clinical data include age, sex, body mass index (BMI), NYHA functional class, smoking status, aetiology (ischaemic vs non-ischaemic) and comorbidities (atrial fibrillation, hypertension, diabetes, chronic kidney disease). Electrocardiographic parameters include QRS duration/morphology, PR interval and atrial fibrillation rhythm. Echocardiographic measurements include LVEF, left ventricular end-diastolic diameter, left ventricular end-systolic volume (LVESV), left atrial diameter and mitral regurgitation severity. Laboratory values include serum creatinine (for estimated glomerular filtration rate calculation), N-terminal pro-B-type natriuretic peptide and haemoglobin. Treatment details comprise CRT device type (CRTP/CRTD) and medication profiles, all predictors used in model development.

Outcome definition

A composite response endpoint is assessed at 6-month post-implantation follow-up (outpatient visits and hospitalisation records). Responders (patients experiencing the positive outcome) are defined as patients meeting ≥1 of the following criteria: (1) NYHA functional class improvement ≥1; (2) LVEF absolute increase ≥5%; or (3) LVESV relative reduction ≥15%—aligned with outcome definitions in meta-analysis-included studies. Non-responders include patients failing to meet any criteria, those with no significant change, or measurements potentially attributable to measurement error.

Missing data handling

Missing data patterns will be assessed initially, with participants exhibiting >30% missingness excluded a priori. Remaining missing values will be imputed using K-Nearest Neighbour (KNN) algorithm; following data normalisation, the k most similar neighbours will be identified based on Gower distance (k optimised via cross-validation, typically 5–10), with continuous and categorical variables imputed using weighted means or modes, respectively.

To evaluate robustness to missing data methodology, KNN will be compared with multiple imputation by chained equations (MICE). For MICE, 5–10 datasets will be generated using predictive mean matching (continuous) and logistic regression (categorical), with all candidate predictors included under the missing-at-random assumption. Model performance will be assessed per dataset, with discrimination and calibration pooled via Rubin’s rules. Concordance between approaches will be evaluated; substantial discrepancies will indicate methodological sensitivity and will be addressed in the Discussion.

Model performance comparison and selection

All nested models will be evaluated concurrently, with discrimination (area under the receiver operating characteristic curve (AUC)) as the primary performance metric. AUC trajectories will be plotted against increasing variable counts to identify the performance plateau—the inflection point beyond which additional variables fail to significantly improve AUC. Adjacent models within the plateau phase will be compared pairwise using DeLong’s test. In instances where AUCs are not statistically significantly different (p≥0.05), the principle of clinical parsimony will guide selection of the more streamlined, clinically practical model as the final recommendation. The final model’s discrimination, calibration and clinical utility will be comprehensively reported.

Discrimination assessment

Discrimination will be evaluated via receiver operating characteristic curve analysis. The AUC will quantify model performance, with values approaching 1.0 indicating superior accuracy (0.5–0.7, low; 0.7–0.9, moderate; >0.9, high). The optimal probability threshold will be identified by maximising Youden’s index. Classification metrics including sensitivity, specificity, positive/negative predictive values, accuracy and F1-score will be comprehensively reported.

Calibration assessment

Calibration will be assessed graphically through calibration plots comparing predicted vs observed probabilities, and quantitatively using the Hosmer-Lemeshow goodness-of-fit test. Patients will be stratified into 10 deciles based on predicted risk; calibration will be tested via χ² analysis. A p value >0.05 indicates adequate model calibration.

Clinical utility assessment

Clinical utility will be evaluated using decision curve analysis (DCA) to quantify net clinical benefit across threshold probability ranges. The DCA will compare three strategies: using the prediction model, treating all patients (treat-all) and treating none (treat-none). Net benefit will be calculated by weighting true-positive benefits against false-positive harms, thereby identifying the probability range where the model provides superior clinical value.

Risk stratification

The final prediction model score will serve as the sole predictor in a logistic regression model to estimate CRT response probabilities and generate a clinical reference table displaying predicted non-response rates across total score ranges. Patients will be stratified into three risk groups based on tertiles of Youden’s J statistic to facilitate rapid clinical assessment.

Statistical methods

Statistical analyses will be conducted throughout meta-analysis, model development and validation. Heterogeneity testing will employ a significance threshold of p<0.10; all other hypothesis tests will be two-sided with statistical significance defined as p<0.05.

Meta-analysis phase

Random-effects models (DerSimonian-Laird) serve as the primary method for pooling effect estimates to incorporate between-study heterogeneity, with fixed-effects models (inverse-variance) employed in sensitivity analyses to assess robustness. Heterogeneity is quantified using I² statistics (I² >50% indicating moderate-to-high heterogeneity), supplemented by Cochrane’s Q test. When substantial heterogeneity is present, subgroup analyses (stratified by study design, population characteristics and outcome definitions) explore potential sources. Publication bias will be evaluated using Egger’s regression test.

Model development and validation phase

Continuous variables are assessed for normality via Shapiro-Wilk tests and Q-Q plots, with normally distributed data presented as mean±SD and non-normal data as median (IQR). Between-group comparisons use independent t-tests (assuming equal variances) or Wilcoxon rank-sum tests otherwise. Categorical variables were reported as counts (percentages) and compared using Pearson’s χ² or Fisher’s exact tests as appropriate.

Model discrimination is quantified using AUCs with 95% CI, with pairwise comparisons performed via DeLong’s test. Calibration is evaluated graphically through calibration plots and quantitatively using the Hosmer-Lemeshow goodness-of-fit test (p>0.05 indicating adequate calibration). Clinical utility is assessed using decision curve analysis (DCA) to compare net benefit across threshold probabilities against treat-all and treat-none strategies.

Statistical software

Meta-analysis phase

EndNote 9.1 facilitates reference management, including deduplication and screening workflow. Microsoft Excel 2021 supports standardised data extraction template development and preliminary data verification. Stata V.17.0 conducts all meta-analytic computations: effect size pooling via random-effects models (DerSimonian-Laird), heterogeneity quantification (I² statistics and Cochran’s Q test), subgroup analysis, sensitivity analysis (including leave-one-out, exclusion of high-risk studies, fixed-effects models and OR-to-RR conversion), meta-regression (REML estimation), publication bias assessment (Egger’s test with funnel plots) and forest plot generation. Certainty of evidence for each pooled predictor will be assessed using the GRADE framework, with summary tables generated in GRADEpro Guideline Development Tool.

Model development and validation phase

All analyses are performed in R (≥4.30.0). Missing values (<30%) are imputed using KNN (VIM package), and the robustness of missing data handling will be assessed by comparison with MICE using the mice package. Risk scores for nested models (A, B, C…) will be calculated via dedicated scripts. Performance evaluation will include the following: discrimination assessed by AUC with 95% CIs using the pROC package, pairwise model comparisons using DeLong’s test, calibration assessed via calibration plots and Hosmer-Lemeshow goodness-of-fit test using the rms package, and clinical utility evaluated via DCA using the dcurves package.