# LDL-C Task Card Task ids: `ldl`, `ldl_nmr` Evidence status: grounded Evidence reviewed: 2026-08-15 ## Target Low-density lipoprotein cholesterol. Continuous-scalar regression. The default research stage is `00_00_visit`, but the Tasks can be configured for another available stage. PhenoBench currently exposes two measurement routes: Implementation facts: For `ldl` (clinical blood test): - target field: `bt__ldl_cholesterol` - target dataset: `blood_tests` - target source: `weizmann` - target unit: mg/dL - valid range: 20.0-500.0 mg/dL - primary metric: R2 For `ldl_nmr` (Nightingale NMR): - target field: `Clinical_LDL_C` - target dataset: `nightingale` - target source: None (Nightingale's primary source is set in the loader) - target unit: **mmol/L**, not "NMR units" - an earlier version of this card hedged the unit, and the hedge is what makes a threshold slip through - valid range: 0.5-8.0 mmol/L - primary metric: R2 Both default to `00_00_visit` with age/sex/BMI covariates at the configured research stage. **The two ids are different measurements in different units on different platforms, not two spellings of one target.** ## HPP data source Target loads from the pinned Weizmann `blood_tests` source. The live operational source HMO table was audited as a candidate replacement. **Audit decision: keep Weizmann for now.** LDL is the one blood target where operational source materially expands coverage (+51.0% net N), but it still replaces ~27% of the current cohort and carries a p99 |Δ| ≈ 32 mg/dL with a small negative bias. A switch should be reconsidered after pre-specifying and meeting an LDL agreement tolerance and validating that the operational source-only participants are genuine baseline draws. Evidence and reconsideration triggers: blood-test source audit §9. ## Why This Matters LDL-C is the canonical causal lipid risk factor for atherosclerotic cardiovascular disease and the main target of lipid-lowering therapy, supported by Mendelian-randomization and treatment-trial evidence. It is clinically familiar, but it is also strongly influenced by genetics and medications, so demographic-only prediction should be weak. The paired clinical/NMR tasks are useful for assay-robustness checks. A result that appears only on one route may reflect assay/platform quirks rather than general LDL biology. **What a row here does not support.** A predicted LDL-C is a same-visit association. It is not a cardiovascular risk estimate and not a treatment indication, and the causal evidence that makes LDL-C important - Mendelian randomisation and lipid-lowering trials - is evidence about *lowering* LDL-C, not about predicting it from other modalities. ## Related HPP Measurements Demonstrated in HPP and persisted in this repository, from an internal review record, test split: - V1 CGM + demographics reaches test R2 = 0.022 (95% CI [-0.020, 0.062]) on a cohort of 3,153. - V1 DXA + demographics reaches R2 = 0.016 (95% CI [-0.024, 0.047]) on 1,892. **Both intervals cross zero, and so do the paired margins against the demographic floor**: DXA **ΔR2 = +0.004 (95% CI [-0.013, +0.023])**, CGM **-0.014 ([-0.050, +0.022])**. On this cohort neither glucose dynamics nor body composition demonstrably predicts LDL-C beyond age, sex and BMI - the expected result for a trait dominated by genetics and lipid-lowering medication. `tg` on the same two modalities clears the floor by +0.122 and +0.053. Those are different cohorts, so read it as an unmatched corpus comparison rather than a controlled contrast. Proposed rather than shown: that genetic scores or dietary features recover what DXA and CGM cannot. Nothing committed answers it. **These are `reviewable` rows, not `final` ones.** Measured across the committed leaderboards: all 787 rows carry `curation_state: reviewable` and none is `final`. A complete fresh v0 re-run is committed to, so every figure in this section is pre-re-run evidence that the re-run supersedes - persisted and citable today, not a frozen result. **Which curation states the Figure 1 export accepts is the export track's live question, not this card's**; read the two counts above, not a gate. ## Clinically Meaningful Variants **LDL-C thresholds are treatment targets stratified by cardiovascular risk, not diagnostic cutoffs, so there is no single boundary to bin on** - and on the NMR route there is a second, mechanical block on top of that. - **The two ids need different conversions, and the factor is analyte-specific.** Converting a cholesterol value between mmol/L and mg/dL uses the molar mass of cholesterol (≈38.7 mg/dL per mmol/L). The triglyceride factor is different (≈88.6, see `tg`). So "convert the NMR route to clinical units" is not one rule applied to a lipid panel; it is per analyte, and using the wrong one produces a plausible-looking number. - **Conversion is not equivalence.** A Nightingale `Clinical_LDL_C` in converted mg/dL is still not the enzymatic assay a guideline threshold was validated on, and the cohorts differ too: the clinical route carries 3,153 participants here against the NMR route's own eligibility. - **A threshold would also cross a medication boundary.** Participants on statins sit below a target *because* of treatment, so a bin mixes untreated biology with treated biology unless medication is modelled, which this Task does not do. The continuous targets stay primary on both routes. `hyperlipidemia_status` is the supported categorical lipid route, and it takes its label from the curated phenotype rather than from a threshold on either field here. ## Baselines And Ceilings The required floor is a no-modality-feature baseline: `NoFeaturesPredictor` contributes no modality features, and `DemographicFloorStrategy` fits age, sex, and BMI. General adult-cohort expectations are low, roughly R2 0.02-0.08 for clinical LDL-C. The task code records surrogate-run context around R2 0.03 for clinical LDL-C and about R2 0.006 for NMR LDL-C. Natural comparators are ApoB, non-HDL-C, triglycerides/remnant cholesterol, and engineered lipidomics/Nightingale feature models. ApoB is the better particle burden target when LDL-C and particle count are discordant. ## Physiological Time-Series V1 the internal evaluation configuration evaluates a residual 1-D CNN trained end-to-end on target-blind samples from the baseline 10-second, 12-lead ECG against same-visit `ldl_nmr`. This is an exploratory representation probe: no direct raw-ECG-to-LDL precedent was found in the research review, and no Nightingale field is allowed in the model input. Interpret held-out delta-R2 over age/sex/BMI as association evidence, not lipid-panel replacement. Held-out exploratory result (test n=884): temporal R² 0.0360 versus tuned age/sex/BMI R² 0.0246, delta-R² +0.0114 (95% CI [-0.0121, +0.0354]); temporal MAE 0.4892 versus 0.4964 mmol/L, delta-MAE -0.0072 mmol/L (95% CI [-0.0152, +0.0010]). Both paired intervals include zero, so this is not evidence of improvement. ## Caveats - Statins and other lipid-lowering therapies can dominate LDL-C levels. - Fasting status matters less than for triglycerides but source/assay route still matters. - Same-panel Nightingale features can leak the NMR LDL target. - LDL-C prediction is not the same as cardiovascular event prediction. - Clinical LDL-C and NMR LDL-C should not be pooled without assay calibration. ## Metrics Primary metric is R2. Secondary MAE should be reported in the task's native units. For shared-card interpretation, report clinical and NMR rows separately and compare demographic-floor delta R2 on matched support when possible. ## References - Source material: `docs/task_cards/source_material/foundation-model-surrogate-targets-v1.md` - Holmes et al., 2015, Mendelian randomization of blood lipids for coronary heart disease: - Sniderman et al., 2024, lipid discordance and apoB context: - Reicher et al. (2025), deep phenotyping of the health/disease continuum, *Nature Medicine*, . Local copy in **HPP documentation**: `assets/papers/hpp/reicher_2025_deep_phenotyping_of_healthdisease_contin/`. ## Benchmark-Track Evidence Benchmark-Track Evidence reviewed: 2026-07-29. V1 DXA and CGM are proposed for both clinical and NMR LDL routes beyond age/sex/BMI. Every clinical/NMR/lipidomics sibling and derived lipid score remains excluded to prevent same-assay reconstruction. See `04-cardiovascular-lipid.md`. ## v0 Final-Test Status The classical release is complete: Ridge/logistic and GBDT are remotely verified on every approved exact cohort and rendered in the Figure 1 task deep dives. ## Open Questions - Which medication fields should be required for statin sensitivity analyses? - Should `ldl` and `ldl_nmr` have a formal assay-concordance report before being compared side by side? - Is ApoB the preferred headline lipid task for Q2 demos, leaving LDL-C as a comparator?