A Temporal Validation Study of Diagnostic Prediction Models for the Screening of Elevated Low-Density and Non-High-Density Lipoprotein Cholesterol

TL;DR

This study evaluates two cholesterol prediction models in young adults, finding one has modest usefulness for non-invasive screening.

Contribution

The study provides a temporal validation of diagnostic models for elevated LDL-C and non-HDL-C in a young adult population.

Findings

The non-HDL-C model showed fair discrimination and 20% investigation reduction at threshold probability.

The LDL-C model had poor discrimination and miscalibration, with negligible investigation reduction.

Updated models showed improved fairness compared to original models.

Abstract

Background/Objectives: Limited accessibility to hypercholesterolemia diagnosis hinders the primary prevention of cardiovascular disease. Therefore, we conducted a prospective, temporal validation study of two diagnostic prediction models, targeting endpoints of elevated low-density lipoprotein cholesterol (LDL-C, ≥160 mg/dL) and non-high-density lipoprotein cholesterol (non-HDL-C, ≥160 mg/dL). Methods: We prospectively recruited workers aged 20–40 years from a single-center, university hospital from March to June 2024 (n = 1099). We determined two diagnostic endpoints: elevated LDL-C and non-HDL-C. The predicted probabilities were derived from the binary logistic regression based on gender, metabolic age, and diastolic blood pressure. We assessed three prediction performances: discrimination from area under the receiver-operating characteristic curve (AuROC); calibration slope (C-slope) and calibration-in-the-large (CITL) from the calibration plot; clinical net benefit from decision curve analysis. Recalibration was based on C-slope and CITL, with a socioeconomic subgroup fairness assessment of AuROC, C-slope, and CITL. Results: From 1099 eligible participants, we identified 135 (12.3%) elevated LDL-C and 251 (22.8%) elevated non-HDL-C cases. The LDL-C model had poor discrimination (AuROC 0.59; 95%-CI, 0.56–0.62), miscalibration (C-slope 0.64; 95%-CI, 0.39–0.88 and CITL −0.14; 95%-CI, −0.27–−0.02), and negligible investigation reduction. The non-HDL-C model had fair discrimination (AuROC 0.67; 95%-CI, 0.64–0.69), miscalibration (C-slope 0.71; 95%-CI, 0.59–0.83 and CITL −0.07; 95%-CI, −0.17–0.03), and 20% investigation reduction at prevalence threshold probability. Updated model fairness improved compared to the original models. Conclusions: Temporal validation demonstrated modest replicability for the elevated non-HDL-C model, with a potential limitation in participants with normal BMI but low muscle and high fat mass. Health practitioners may use updated elevated non-HDL-C models as a non-invasive triage strategy in young adults, with threshold probabilities within the positive clinical net benefit ranges. Further external validation studies in a larger and more diverse population are necessary.

Figures