When LDL-C Equations Disagree: A Zero-Cost Uncertainty Signal and an Interpretable Calibration Fix
This study asks whether the disagreement among three widely available LDL-C estimation equations (Friedewald, Sampson/NIH, Martin-Hopkins) can serve as a free uncertainty signal at clinical decision thresholds, and whether a simple interpretable calibration model can rescue misclassification in the disagreement subgroup. Using 10,799 All of Us lipid panels with direct LDL-C as the reference standard, panels were labeled Agree when all three equations fell on the same side of 70, 100, or 130 mg/dL and Disagree otherwise. Agreement was common (86–92% of panels) with 92–96% accuracy, while disagreement (8–14%) dropped accuracy to 48–61%. A regime-aware calibration model achieved a mean absolute error of 8.98 mg/dL, statistically indistinguishable from the best ML ensemble (9.01 mg/dL; 95% CI for the difference −0.35 to 0.29 mg/dL). In 14,549 external MIMIC-IV panels, the model beat the best individual equation by 3.5–9.0 percentage points and majority vote by 18.5–25.2 percentage points among split-threshold panels; calibration improved accuracy by 19–25 percentage points, and hybrid routing reached 90–93% internally and 90–94% externally. The authors conclude that equation disagreement is a zero-cost uncertainty signal and that a simple interpretable model can match the best ML ensemble while improving classification.