# Triglycerides Task Card Task ids: `tg`, `tg_nmr` Evidence status: grounded Evidence reviewed: 2026-08-15 ## Target Fasting or same-visit triglyceride concentration. 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 `tg` (clinical blood test): - target field: `bt__triglycerides` - target dataset: `blood_tests` - target source: `weizmann` - target unit: mg/dL - valid range: 20.0-1500.0 mg/dL - primary metric: R2 For `tg_nmr` (Nightingale NMR): - target field: `Total_TG` - 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: 0.2-15.0 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.** Rank agreement is high (ρ 0.990), but operational source + `events` yields net −14.7% baseline coverage with ~33% cohort turnover and the largest absolute tail (p99 |Δ| ≈ 45 mg/dL). Rationale and reconsideration triggers: blood-test source audit §9. ## Why This Matters Triglycerides reflect triglyceride-rich lipoproteins, remnant metabolism, diet, insulin resistance, sleep, alcohol, and activity. They are more lifestyle responsive and noisier than LDL-C, making them useful for testing whether multimodal representations capture dynamic metabolic state. The clinical and NMR routes should be treated as paired assay views, not as interchangeable labels. **What a row here does not support.** A predicted triglyceride value is a same-visit association, not a cardiovascular risk estimate and not a dietary or lifestyle recommendation. Triglycerides also move with the fasting state, so a row is a statement about this cohort's draw conditions as much as about biology. ## Related HPP Measurements Demonstrated in HPP and persisted in this repository, from an internal review record, test split: - V1 DXA + demographics reaches test R2 = 0.210 (95% CI [0.153, 0.262]) on a cohort of 3,573. - V1 CGM + demographics reaches R2 = 0.151 (95% CI [0.102, 0.195]) on 5,752. **These are the largest lipid-family effects in this card set**, though that is a cross-cohort observation rather than a tested ranking. The contrast with `ldl` on the same two modalities is the point: LDL-C sits at 0.016 and 0.022, while triglycerides are clearly predictable from body composition and glucose dynamics. **These are different cohorts, not one**: TG scores 3,573 and 5,752 participants, LDL 1,892 and 3,153, under four different `cohort_id` hashes, and raw R2 across two targets are different estimands besides. So the lifestyle-responsive versus genetically-driven reading stays a **hypothesis consistent with an unmatched corpus comparison**, pending a common-support evaluation - not something these rows establish. **Both TG tracks clear the demographic floor on the paired comparison**, which is the quantity a modality claim rests on: DXA **ΔR2 = +0.122 (95% CI [+0.086, +0.166])** and CGM **+0.053 ([+0.017, +0.086])**. `ldl` on the same two modalities does not clear its own floor (+0.004 and -0.014, both spanning zero). Each of those is a within-track statement about its own cohort. Read together they are consistent with the lifestyle-versus-genetic split, and they do not test it: TG and LDL-C are different targets on different participants, so no paired comparison exists here. Related, in the wider HPP literature: Shilo et al. analysed the TG/HDL ratio alongside other insulin-resistance surrogates in 10,114 non-diabetic HPP adults and found the surrogates only moderately inter-correlated, capturing distinct metabolic facets (Shilo et al. 2026). Triglycerides feed two of those surrogates and the `tyg` Task, so a result here is entangled with that family by construction. **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 **Triglyceride thresholds exist and are the most transferable in this family, and three things still block a bin on these fields.** - **Units, with a factor different from the other lipids.** Converting triglycerides between mmol/L and mg/dL uses ≈88.6 mg/dL per mmol/L, against ≈38.7 for cholesterol (see `ldl`). A single "NMR to clinical" rule applied across a lipid panel is wrong for at least one analyte, and wrong quietly. - **The fasting state is not verified.** Clinical triglyceride cutoffs are stated for a defined fasting condition, and non-fasting values run higher. This Task's own title says "fasting or same-visit", and nothing in the Task checks which - the same unresolved fasting-metadata question `fpg` records. - **Platform.** A Nightingale `Total_TG` is not the enzymatic assay the cutoffs were set on, and converting does not make it one. The continuous targets stay primary on both routes. `hyperlipidemia_status` is the supported categorical lipid 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 demographic floor R2 around 0.11-0.12. Natural comparators include ApoB, LDL-C, TyG, fasting glucose, CGM summary features, dietary features, sleep/OSA summaries, and engineered lipidomics or NMR feature models. Same-assay lipid features are a high leakage risk. ## Caveats - Triglycerides are highly variable and sensitive to fasting compliance, recent diet, alcohol, illness, and medication. - CGM or diet features may legitimately predict TG via insulin resistance and carbohydrate exposure, but same-window interpretation should remain current-state, not prospective. - NMR TG and clinical TG should not be merged without assay calibration. - Strong TG prediction does not by itself imply remnant-cholesterol or MACE prediction. ## Metrics Primary metric is R2. Secondary MAE should be reported in native units. Useful audits include log-scale sensitivity, demographic-floor delta R2, matched support across `tg`/`tg_nmr`, and medication/fasting-status sensitivity when available. ## References - Source material: `docs/task_cards/source_material/foundation-model-surrogate-targets-v1.md` - Holmes et al., 2015, triglycerides in Mendelian-randomization CHD analysis: - Zhao et al., 2024, remnant cholesterol/triglycerides and cardiometabolic multimorbidity: - Shilo et al. (2026), heterogeneity of insulin-resistance surrogates, *medRxiv*, . Local copy in **HPP documentation**: `assets/papers/hpp/shilo_2026_heterogeneity_of_insulin_resistance_surr/`. ## Benchmark-Track Evidence Benchmark-Track Evidence reviewed: 2026-07-29. V1 DXA and CGM are proposed for both clinical and NMR triglyceride routes beyond age/sex/BMI. Every clinical/NMR/lipidomics sibling and lipid-derived 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 - Should the primary metric be computed on raw TG or log-transformed TG for heavy-tail robustness? - What fasting-status evidence is available for the HPP blood draw used here? - Which same-assay lipid features should be excluded from predictor allowlists?