TUDOR, an ascertainment-aware model for LDLR genetic-testing triage in specialist lipid clinics: Development, validation, and UK Biobank transport.

BACKGROUND

Familial hypercholesterolemia remains underidentified despite a clear causal pathway, effective treatment, and efficient cascade testing after genetic confirmation. Contemporary lipid clinics increasingly evaluate treated patients, cascade-screened relatives, and people with mixed metabolic dyslipidemia, in whom the recorded low-density lipoprotein cholesterol (LDL-C) phenotype may no longer resemble the untreated proband phenotype assumed by conventional criteria. TUDOR was developed as an ascertainment-aware model for prioritizing confirmatory LDLR genetic testing in specialist lipid clinics.

OBJECTIVE

The primary objective of this study was to determine whether TUDOR improves discrimination for genetically confirmed LDLR-mediated FH in specialist lipid-clinic triage compared with available-data implementations of established criteria.

METHODS

TUDOR is an 11-feature elastic-net logistic model estimating genetically confirmed LDLR carrier status from routine lipid-clinic variables. It reconstructs untreated LDL-C, separates receptor-mediated LDL signal from mixed metabolic dyslipidemia using triglycerides, non-high-density lipoprotein cholesterol, and a type-2-diabetes-by-LDL interaction, and encodes index-referral vs cascade-relative ascertainment. Validation used bidirectional geographic internal-external validation in Wales, a complete-case head-to-head comparison against electronic Dutch Lipid Clinic Network (eDLCN), familial hypercholesterolaemia case ascertainment tool (FAMCAT), Simon Broome, and Make Early Diagnosis to Prevent Early Deaths (MEDPED), UK Biobank (UKB) lipid-clinic-eligible frozen transport, local coefficient updating, whole-UKB stress testing, LDL-C reconstruction validation, and subgroup analyses.

RESULTS

In the Welsh complete-case head-to-head cohort (n = 1274; 311 carriers), TUDOR reached an area under the receiver operating characteristic curve (AUC) of 0.760 compared with 0.652 for eDLCN, 0.600 for FAMCAT, 0.569 for Simon Broome, and 0.524 for MEDPED. Geographic internal-external validation yielded AUCs of 0.732 (95% CI 0.707-0.757) and 0.770 (95% CI 0.739-0.797), with a pooled frozen Welsh AUC of 0.746 (95% CI 0.726-0.764). Frozen transport to the UKB lipid-clinic-eligible cohort gave an AUC of 0.669 (95% CI 0.650-0.687); local coefficient updating gave an apparent AUC of 0.756 (95% CI 0.734-0.777). Whole-UKB stress testing retained signal at an AUC of 0.631 (95% CI 0.622-0.641). Back-calculation was validated in 649 Welsh patients with measured pretreatment LDL-C (mean absolute error (MAE) 1.20 mmol/L, 95% CI 1.13-1.28).

CONCLUSION

TUDOR improved rank-ordering for genetically confirmed LDLR carrier status in Welsh specialist lipid-clinic validation and retained external signal when transported frozen to a UKB lipid-clinic-eligible cohort. Local coefficient updating recovered Welsh-range apparent discrimination, supporting transportability of the feature architecture but not universal portability of the frozen Welsh equation. TUDOR is a triage aid for confirmatory genetic testing, not a diagnostic replacement or population screener; prospective deployment requires local calibration, threshold-specific carrier yield, decision-curve analysis, and safety monitoring.