RARE-CROSS: development and evaluation of an operational framework for trustworthy rare-event modelling in cross- sectional health surveys

Abstract Background Rare-event analyses in cross-sectional health surveys are often treated primarily as class-imbalance problems, although validity can be compromised by outcome definition, temporal ambiguity, detection-sensitive predictors, complex sampling, missingness, data leakage, probability distortion after balancing and poor transportability. Existing standards address many of these threats separately. We developed RARE-CROSS to integrate them into an operational and auditable workflow. Methods RARE-CROSS comprises nine domains organised around DEFINE, PROTECT and PROVE. Development mapped recurrent failure modes to established guidance and translated them into operational guardrails, minimum reporting items and enhanced options. The framework was cross-walked against TRIPOD + AI, PROBAST + AI, PRICSSA, simulation-study and external-validation guidance. Targeted Monte Carlo experiments examined prevalence effects, SMOTE leakage, altered class priors, detection-related predictors and outcome-dependent missingness. An empirical demonstration used 52,572 adults, 2,778 self-reported cardiovascular disease events and 11 African WHO STEPS survey-country datasets from 2014 to 2019. Results Mean ROC AUC changed little as simulated prevalence varied from 1% to 10% (0.793 to 0.773), whereas PR AUC increased from 0.105 to 0.328. Correctly nested SMOTE produced PR AUC 0.166, but pre-split SMOTE produced an artefactual 0.799. Prior-prevalence correction reduced the Brier score from 0.178 to 0.023 without changing ranking. A detection-related covariate increased PR AUC from 0.224 to 0.327. In the STEPS demonstration, pooled leave-one-country-out PR AUC was 0.079, with substantial country heterogeneity. Conclusions RARE-CROSS provides an evaluated, algorithm-agnostic architecture for designing, auditing and reporting rare-event analyses in cross-sectional surveys. It integrates prevalence-aware performance, calibration, leakage-safe preprocessing, missingness structure, detection-dependence stress testing, survey design, uncertainty and performance beyond development data. The framework is intended to complement, not replace, established reporting and risk-of-bias standards.