# ApoB Task Card Task id: `apob` Evidence status: grounded Evidence reviewed: 2026-08-15 ## Target Apolipoprotein B concentration from the HPP Nightingale NMR panel. Continuous-scalar regression. The default research stage is `00_00_visit`, but the Task can be configured for another available stage. Implementation facts: - target field: `ApoB` - target dataset: `nightingale` - target source: None - default research stage: `00_00_visit` - valid range: 0.3-3.0 g/L - primary metric: R2 ## Why This Matters ApoB approximates the number of atherogenic lipoprotein particles. It is often more clinically informative than LDL-C when LDL-C and particle number are discordant, especially in insulin resistance or hypertriglyceridemia. For PhenoBench, ApoB is a strong lipid-axis target because the demographic floor is low and the clinically meaningful signal should come from molecular, dietary, metabolic, or genetic information rather than age/sex/BMI alone. **What a row here does not support.** A predicted ApoB is a same-visit association with an NMR panel value. It is not a cardiovascular risk estimate, not a treatment indication, and not interchangeable with a clinical ApoB immunoassay result, which is what every published ApoB threshold is defined on. ## Related HPP Measurements Demonstrated in HPP and persisted in this repository, from the deep dive in an internal review record, test split: - The best raw scores come from GBDT: V1 CGM + demographics reaches test R2 = 0.075 (95% CI [0.046, 0.106]) on a cohort of 9,137; V1 DXA + demographics reaches R2 = 0.029 (95% CI [-0.006, 0.059]) on 6,574. - On the Ridge paired rows, CGM reaches **ΔR2 = +0.033 (95% CI [+0.014, +0.054])** and clears the demographic floor; DXA reaches **+0.003 ([-0.010, +0.017])** and does not. - The GBDT paired deltas do not show the same split: CGM reaches +0.017 ([-0.009, +0.043]) and DXA reaches -0.008 ([-0.031, +0.013]); both intervals cross zero. Added CGM signal is therefore method-specific: Ridge demonstrates it here; GBDT does not. Body composition adds nothing demonstrable with either method. The contrast with `tg` survives and is better stated on the deltas: on DXA, triglycerides clear the floor by +0.122 ([+0.086, +0.166]) where ApoB manages +0.003 on Ridge. Those are different cohorts, so it is an unmatched corpus comparison. Proposed rather than shown: that dietary, genetic or learned representations recover ApoB where DXA does not. **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 **ApoB is the one lipid target here with genuinely actionable thresholds, and they still do not transfer to this measurement.** Guideline ApoB goals are stated in mg/dL against risk stratum and are defined on standardised clinical immunoassays. This Task's target is the **Nightingale NMR `ApoB` in g/L**. Two independent mismatches block a bin: - **Units.** g/L and mg/dL differ by a factor of 100, so a threshold applied without conversion is wrong by two orders of magnitude - the failure mode is silent, because both numbers look plausible in their own scale. - **Platform.** Even converted, an NMR-derived ApoB is not the immunoassay the goals were validated against. Agreement between the two is an empirical question this repository has not measured, and until it is measured the conversion is arithmetic rather than equivalence. A risk-stratified target is also not a diagnostic cutoff: ApoB goals depend on the patient's cardiovascular risk category, so a single-threshold binary would misrepresent the guideline even on the right assay in the right units. The continuous target stays primary. If a categorical lipid Task is wanted, the curated `hyperlipidemia_status` label is the supported route. ## Baselines And Ceilings The required floor is a no-modality-feature baseline: `NoFeaturesPredictor` contributes no modality features, and `DemographicFloorStrategy` fits age, sex, and BMI. The task code records surrogate floor R2 around 0.02. Natural comparators include LDL-C, non-HDL-C, triglycerides, and engineered Nightingale/lipidomics feature models. A model that only improves LDL-C but not ApoB should not be claimed to capture atherogenic particle burden broadly. ## Caveats - ApoB can be medication-confounded, especially by statins and other lipid-lowering therapy. - NMR ApoB is not the same assay as a clinical immunoassay; assay agreement should be checked before making clinical-equivalence claims. - Same-visit lipid features may be direct or near-direct measurement leakage. - A good ApoB score supports lipid-axis information, not prospective MACE prediction unless paired with incident outcomes. ## Metrics Primary metric is R2. Secondary MAE is useful in g/L. Audit views should include demographic-floor delta R2, split Ns, and medication/statin sensitivity when available. ## References - Source material: `docs/task_cards/source_material/foundation-model-surrogate-targets-v1.md` - Thanassoulis et al., 2014, statin-trial meta-analysis relating apoB change to risk reduction: - Sniderman et al., 2024, apoB discordance and cardiovascular prevention: - Han et al. (2026), multi-omics dietary patterns and cardiometabolic risk, *medRxiv*, . Local copy in **HPP documentation**: `assets/papers/hpp/han_2026_multi_omics_characterization_of_biologic/`. ## Benchmark-Track Evidence Benchmark-Track Evidence reviewed: 2026-07-29. literature and HPP review prioritizes V1 DXA and CGM beyond age/sex/BMI. All lipid/Nightingale inputs remain excluded because same-assay siblings can reconstruct ApoB. 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 - Should clinical ApoB, if available in HPP blood chemistry, become a separate task or a validation check for NMR ApoB? - What medication exclusion or stratification policy should be used for lipid-lowering therapy? - Which lipidomics feature set should define the near-term engineered ceiling?