When three LDL-C equations disagree at thresholds, accuracy falls to 48%-61%, and an interpretable calibration model raises classification accuracy by 19-25 percentage points

Disagreement among the equations is itself a zero-cost uncertainty signal that identifies patients at highest misclassification risk. Whereas laboratories typically report only one of the three computable equations, this study uses whether all three fall on the same side of a threshold as a stratification variable, yielding risk stratification without additional testing. Based on 10,799 All of Us lipid panels with direct LDL-C as the reference standard, stratified at the 70, 100, and 130 mg/dL thresholds; the agreement group comprised 86%-92% of panels with 92%-96% accuracy, while the disagreement group comprised 8%-14% with 48%-61% accuracy.

A simple, interpretable regime-aware calibration model improves threshold classification without sacrificing accuracy. The model was compared with 7 machine learning methods under 5-fold cross-validation, achieving a mean absolute error of 8.98 mg/dL that matched the best ML ensemble at 9.01 mg/dL (95% CI for the difference, -0.35 to 0.29 mg/dL) while remaining interpretable. Internal 5-fold cross-validation mean absolute error comparison with confidence interval; external validation on 14,549 hospital-based MIMIC-IV panels.

The classification benefit of calibration is largest in the high-risk split-threshold subgroup. External validation showed the model outperformed the best-performing individual equation at each threshold by 3.5-9.0 percentage points and the majority vote by 18.5-25.2 percentage points; calibration improved accuracy by 19-25 percentage points. External validation cohort of 14,549 MIMIC-IV panels; hybrid routing achieved 90%-93% accuracy internally and 90%-94% externally.

The result is intended for laboratory and clinical settings that use standard lipid panels with direct LDL-C as the reference: any institution able to compute the Friedewald, Sampson/NIH, and Martin-Hopkins equations can first stratify by Agree/Disagree and then apply regime-aware calibration and hybrid routing to the disagreement subgroup. Internal data come from All of Us and external data from hospital-based MIMIC-IV intensive care panels, so the method applies to threshold classification of adult lipid panels, particularly at decision points near 70, 100, and 130 mg/dL.

Only the abstract was loaded, so figures, model details, and subgroup analyses are missing; the specific form of the calibration model, the trigger conditions for hybrid routing, and performance in other populations (beyond All of Us and MIMIC-IV) still require the full text to confirm. In addition, the abstract does not describe how the three equations differ within the disagreement group, nor how the calibration model behaves at extreme lipid values, which are open questions worth watching before application.