Predict LDL without extra testing
A new machine‑learning tool can predict LDL cholesterol from standard lab inputs — promising a lower‑cost way to flag high LDL risk in diabetes and cardiometabolic populations. (medscape.com)
Medscape flagged the finding as an ECIM 2026 presentation and published coverage on March 31, 2026. (medscape.com) A Johns Hopkins team used baseline data from 3,509 ACCORD trial participants to train models that predicted 12‑month LDL‑C change after statin start, reporting Lasso as the top performer with an LDL‑C R² ≈ 0.52 and point‑estimate error around 18.3 mg/dL, and providing individual prediction intervals with ≈90% empirical coverage via conformal prediction. (medrxiv.org) A separate 2024 Laboratory Medicine analysis found a novel two‑step ML model achieved RMSE 7.015 for directly measured LDL‑C versus Friedewald RMSE 12.112, Martin 8.084, and Sampson 8.492, with the ML approach reaching an 85.1% concordance with direct assays. (academic.oup.com) A 2025 Anatolian Journal validation used 25,991 lipid profiles (triglycerides 400–800 mg/dL) to train a stacked random‑forest/XGBoost model (model‑3) that produced balanced accuracy 99.3%, precision 98.9%, recall 98.9%, specificity 99.8%, and a correlation with direct LDL of r=0.996 (mean bias −0.2 mg/dL). (anatolianjmed.org) An explainable‑AI study drawing on 60,217 same‑day lipid panels from Ankara Etlik City Hospital demonstrated XAI techniques (SHAP) can identify which routine inputs drive LDL estimates but also flagged variability from differing direct‑assay methods across labs. (read.qxmd.com) Authors and laboratory‑medicine reviewers say the path forward requires prospective validation across diverse biochemistry platforms, harmonization of assay standards, and technical integration into laboratory information systems before routine clinical deployment. (pmc.ncbi.nlm.nih.gov)
