Accurate LDL cholesterol estimation underpins one of medicine's most consequential clinical decisions — whether and how aggressively to treat cardiovascular risk. The standard Friedewald equation, still used in most labs worldwide, is known to systematically underestimate LDL-C at low concentrations and in patients with high triglycerides, precisely the populations at highest cardiac risk. A more accurate replacement that is also computationally simple enough for routine clinical adoption would represent meaningful progress.
This study, published in JAMA Cardiology, applied multivariate adaptive regression splines (MARS) — a form of machine learning — to the Very Large Database of Lipids, a cross-sectional dataset of clinical lipid measurements spanning 2015 to 2019 from both adult and pediatric patients. The resulting LDL-C-MH-MARS equation was benchmarked against four established estimation methods: the Friedewald, Sampson-NIH, Modified Sampson, and Martin-Hopkins equations. External validation was conducted in two independent reference datasets — one from Mayo Clinic covering broad LDL-C concentrations, and one from the FOURIER trial, which specifically captured patients with very low LDL-C values while on evolocumab therapy. Both validations used preparative ultracentrifugation as the true reference standard, the most direct measurement method available.
The MARS-derived equation matched or exceeded the performance of the Martin-Hopkins equation — currently considered one of the most accurate estimation methods — while achieving meaningful computational simplification. Performance was notably robust in the low LDL-C range, which is exactly where older equations fail and where therapeutic decisions about PCSK9 inhibitors are most consequential.
This finding is clinically relevant but should be interpreted carefully. The underlying database, while large, is a convenience sample rather than a fully representative population cohort. MARS models, even when simplified, require implementation infrastructure that smaller or resource-limited labs may lack. The true test will be prospective validation in diverse clinical settings. Nonetheless, for a field where imprecise LDL estimation has quietly influenced millions of treatment decisions, a validated, simplified improvement to the estimation pipeline is genuinely useful — incremental in method, but potentially meaningful in population-level impact.