Detecting Friedewald-Substituted "Direct" LDL Cholesterol in a National Multi-Site Electronic Health Record Cohort: Prevalence and Benchmarking Impact.

BACKGROUND

When LDL cholesterol (LDL-C) estimation equations are benchmarked against "direct" LDL-C (LDLdirect) values from electronic health records (EHRs), the implicit assumption is that reported direct values derive from chemical assays rather than from calculations. This assumption has not been systematically tested.

METHODS

We analyzed 19 470 same-day lipid panels from the All of Us Research Program (50+ health provider organizations). We defined the effective divisor, Deff = TG/(TC - HDL-C - LDLdirect), where TG is triglycerides and TC is total cholesterol, and tested for evidence that reported "direct" values were Friedewald-calculated (Deff ≈ 5.0), with validation on the National Health and Nutrition Examination Survey (NHANES) and the Medical Information Mart for Intensive Care IV (MIMIC-IV).

RESULTS

Of 19 470 panels (from All of Us), 7.6% (1482) were algebraically confirmed Friedewald-substituted (|LDLFW - LDLdirect| ≤ 0.02 mg/dL [0.0005 mmol/L], where LDLFW is the Friedewald-calculated LDL-C) and 31.1% met a broader criterion (≤ 0.5 mg/dL [0.013 mmol/L]). The Deff distribution showed a 27.6-fold spike at 5.0 (z = 68 vs empirical background; z = 263 vs shuffled-LDL null). Contamination deflated observed Friedewald mean absolute error (MAE) by 30.7% (11.85 vs 17.10 mg/dL [0.306 vs 0.442 mmol/L]; true MAE 44.3% higher). Near treatment thresholds, contamination masked 12 percentage points of misclassification among near-threshold cases at 70 mg/dL [1.81 mmol/L] (26% apparent vs 38% true). Within-patient discordance confirmed substitution (39.6-fold gap, P < 10⁻⁵). NHANES validation showed 28.3-fold enrichment. In MIMIC-IV, the spike was absent in measured panels (0.8×) but present in calculated panels (z = 210).

CONCLUSIONS

A minimum of 7.6% of "direct" LDL-C values in this national multi-site EHR cohort are algebraically confirmed Friedewald-substituted (up to 31.1% under a broader criterion), deflating observed equation error and attenuating the apparent advantage of improved methods. Deff decontamination should be a standard preprocessing step for EHR-based equation benchmarking.