Accurate, Race-Free LDL-C Estimation in Non-Fasting Settings: A Machine-Learning Study in 3,477 Adults

Importance: Traditional LDL-C testing barriers – mandatory 9-12 hour fasting and in-person visits – disproportionately limit access for rural populations (60% of US counties lack cardiology services), shift workers, patients with diabetes, and caregivers. These barriers persist despite 2016 European guidelines endorsing non-fasting lipid panels and explosive post-COVID telehealth growth. Objective: To determine whether machine learning can maintain and enhance LDL-C accuracy in real-world access-expanding scenarios: non-fasting collection, remote assessment without vital signs, and race-free inputs with equity audit. Design, Setting, and Participants: Cross-sectional analysis of the All of Us Research Program (n=3,477 adults; 40.1% tested outside traditional fasting windows). We evaluated accuracy across fasting proxies, labs-only configurations, and racial/ethnic groups using an 80/20 train/test split with bootstrap confidence intervals. Main Outcomes and Measures: Mean absolute error (MAE) and calibration in likely non-fasting states (blood drawn after 10:00 local time, triglycerides (TG) >= 200 mg/dL, glucose > 110 mg/dL); labs-only configuration (no blood pressure); racial equity (Black/African American vs White MAE gap); economic impact of single-visit workflows. Results: Among 696 test participants, 279 (40.1%) were tested in likely non-fasting states. Standard equations degraded substantially in non-fasting conditions (Friedewald MAE 29.05 mg/dL vs 25.87 mg/dL fasting; calibration slope remained far from ideal at 0.58-0.61), while the ML system maintained accuracy (MAE 24.04 mg/dL non-fasting vs 23.18 mg/dL fasting; slope 0.99-1.07) – yielding 17.2% improvement over Friedewald in non-fasting states vs 10.4% when fasting. Removing blood pressure measurements changed MAE by only -0.12 mg/dL (90% CI -0.33 to 0.08), meeting non-inferiority within +/-0.5 mg/dL margin and enabling telehealth workflows. The system achieved racial equity without race input (Black/African American vs White gap -0.19 mg/dL, 95% CI -4.12 to 3.74, indicating no meaningful difference). Economic modeling showed eliminating fasting requirements prevents about 4,000 repeat visits per 10,000 tests. Base-case net savings of $185,000 per 10,000 tests (40% repeat-visit rate at $50/visit minus $15,000 implementation/QA costs). Break-even occurs at 750 tests in year one, then 250 tests/year thereafter. Conclusions and Relevance: Machine learning enables accurate LDL-C assessment without fasting or in-person visits, addressing critical access barriers for rural, diabetic, and underserved populations. With roughly 40% of real-world testing already occurring outside fasting windows, single-visit workflows can substantially reduce repeat visits while maintaining measurement quality and achieving racial equity.