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

ImportanceCurrent LDL-C testing requires 9-12 hour fasting and in-person visits, creating an access crisis: 40% of lipid panels occur outside fasting windows (yielding unreliable results), 60% of US counties lack cardiology services, and millions of patients with diabetes cannot safely fast. Meanwhile, telehealth infrastructure expanded 38-fold post-COVID, yet lipid workflows remain anchored to 1970s protocols. This mismatch drives ~ 20 million unnecessary repeat visits annually, disproportionately burdening Medicaid populations, essential workers, and rural communities. ObjectiveTo demonstrate that machine learning can transform lipid testing from a fasting-dependent, clinic-based bottleneck into an accurate, equitable, telehealth-ready service--eliminating three structural barriers simultaneously: fasting requirements, in-person visits, and racial algorithmic bias. Design, Setting, and ParticipantsCross-sectional analysis of All of Us Research Program (n=3,477; test n=696). Crucially, 40.1% were tested outside traditional fasting windows, reflecting real-world practice. We evaluated performance stratified by fasting status, telehealth feasibility (labs-only configuration), racial equity metrics, and economic impact. Main Outcomes and MeasuresPrimary: MAE and calibration in non-fasting states. Secondary: Labs-only non-inferiority ({+/-}0.5 mg dL-1margin); racial equity (Black-White performance gap); economic savings from eliminated repeat visits; and classification accuracy at treatment thresholds (70, 100, 130 mg dL-1). ResultsThe ML system demonstrated paradoxical superiority in non-fasting conditions--precisely when needed most. While equations deteriorated (Friedewald MAE 29.1 vs 25.9 mg dL-1fasting, slopes 0.58-0.61), ML maintained accuracy (24.0 vs 23.2 mg dL-1, slopes 0.99-1.07), achieving 17.2% improvement over Friedewald when non-fasting vs 10.4% fasting. Labs-only configuration proved non-inferior (MAE=-0.12, p<0.001), enabling national retail-pharmacy and home-testing workflows. The system achieved racial equity without race input (Black-White gap -0.19 mg dL-1, CI includes zero) while providing greatest improvement for Black patients (19% vs 11% for White). Economically, eliminating 4,000 repeat visits per 10,000 tests helps address an estimated $2 billion annual repeat-testing cost burden and yields $815,000 total savings per 10,000 tests ($245,000 direct healthcare, $570,000 patient costs), with break-even at just 750 tests. Conclusions and RelevanceThis ML approach helps address an estimated $2 billion annual problem of repeat testing while tackling three critical quality gaps in cardiovascular prevention: delayed treatment initiation, poor monitoring adherence, and specialty access barriers. By enabling accurate non-fasting, telehealth-compatible, race-free LDL-C estimation, it transforms lipid testing from an access barrier into an access enabler--particularly for the Medicaid, Medicare Advantage, and rural populations who drive both cost and outcomes in value-based care. From a technical standpoint, implementation requires only routine labs and <100 ms computation, making deployment feasible with existing infrastructure.