Cardiovascular disease remains the leading cause of morbidity and mortality in women. Investigators reported this protocol to update the prior systematic review by Baart et al., focusing on models developed for women or those incorporating female-specific predictors. The principal objective is to identify published multivariable prediction models that estimate incident arterial cardiovascular events in women, catalog model predictors and validation methods, and evaluate model performance and clinical usability. A core question driving the review is whether the addition of female-specific variables (for example, pregnancy complications or premature menopause) provides incremental predictive information beyond traditional risk factors.
The review team will update the original search strategy in collaboration with a medical librarian and will search MEDLINE and EMBASE only. The search window covers July 27, 2017 through May 2026. Investigators noted that letters, conference abstracts and presentations will be excluded. Reference lists of included articles will be screened to capture studies potentially missed by the two databases. The protocol has been registered on PROSPERO (CRD420261420766). The authors stated that no additional databases, trial registries or author contact is planned.
Included reports must present at least one multivariable prediction model, tool or score for individual CVD risk estimation in women. The team will accept four study types: models developed in female-only cohorts; general-population models that include sex as a predictor or effect modifier; external validation studies performed in female-only or sex-stratified samples; and incremental value studies that evaluate the addition of female-specific predictors to existing models. Investigators listed candidate female-specific predictors to be considered, including early menopause, polycystic ovary syndrome, pre-eclampsia, gestational diabetes, gestational hypertension, preterm delivery, fetal growth restriction, endometriosis and perinatal depression.
The review will include models predicting fatal and non-fatal arterial cardiovascular outcomes such as coronary heart disease, myocardial infarction, stroke, cardiovascular death and congestive heart failure. Studies restricted to venous disease, cross-sectional validation papers, or models developed exclusively in disease-specific cohorts (for example, only diabetes or chronic kidney disease patients) will be excluded. Studies in female-specific or obstetric populations remain eligible.
Study records will be managed with Covidence. Title/abstract and full-text screening will be performed independently and in duplicate. Data extraction will follow the CHARMS checklist and be completed by one author with independent verification by a second. Where information is not reported in the source article, the protocol specifies data will be recorded as unavailable.
Extracted items will include study design, population characteristics, geographic setting, outcome definitions, prediction horizon, modelling method (statistical, machine learning or AI), number and type of predictors, handling of missing data, internal and external validation procedures, and model performance measures including discrimination, calibration and reclassification indices. The reviewers will also record measures of clinical usability such as number of predictors, whether the model equation is reported, and availability of an online calculator. For incremental value studies, the protocol specifies extraction of which added predictors provided incremental value and the comparator model used.
The protocol prescribes use of the PROBAST+AI tool to assess risk of bias for each included study. Assessments will be completed independently by one reviewer and checked by a second, with disagreements resolved by consensus. The authors indicated that risk of bias judgments will inform interpretation of findings and the overall strength of evidence. No formal certainty-of-evidence framework (for example, GRADE) will be applied; instead, strength will be considered in the context of clinical and methodological heterogeneity plus risk of bias.
The primary synthesis will be descriptive. Quantitative meta-analysis will be considered only when at least three external validation studies evaluate the same prediction model using comparable outcomes, prediction horizons and performance measures. If that threshold is met, the protocol describes pooling c-statistics and calibration measures using random-effects meta-analysis. Investigators plan to transform c-statistics to the logit scale for pooling, derive standard errors from reported data, and back-transform pooled estimates to the original c-statistic scale. Heterogeneity will be assessed with I2 and τ2, and a 95% prediction interval will be reported when possible. Sensitivity analyses based on risk-of-bias assessments will be considered, including exclusion of studies judged to have significant bias. The protocol notes that analyses will be performed in R. The authors do not plan a formal assessment of publication bias because quantitative synthesis is expected only for a limited number of external validations.
The review is deliberately limited to MEDLINE and EMBASE to maintain continuity with the Baart et al. search strategy; the authors acknowledged this constraint may miss eligible studies despite planned reference-list screening. Heterogeneity in study populations, outcome definitions, prediction horizons and reported performance metrics is anticipated and may preclude meta-analysis for many models. The protocol specifies that availability of at least three comparable external validations will be the minimum criterion for pooling, but the presence of three studies alone will not automatically trigger meta-analysis if clinical or methodological differences remain substantial.
A central unresolved question the review aims to address is whether incorporation of female-specific predictors materially improves risk estimation beyond traditional factors; the protocol does not prespecify which specific effect size would constitute meaningful incremental value. The review will record whether incremental value studies report such metrics, but the protocol did not report thresholds for clinical action. Finally, where required methodological details are absent from primary studies, the reviewers will record those items as unavailable, limiting downstream interpretation and synthesis.