# TyG Task Card Task id: `tyg` Evidence status: grounded Evidence reviewed: 2026-08-15 ## Target TyG is the Triglyceride-Glucose index, a fasting-blood insulin-resistance surrogate: ```text TyG = ln(triglycerides_mg_dl * fasting_glucose_mg_dl / 2) ``` Implementation facts: - target field: `None`; this is a derived formula target, computed from `bt__triglycerides` and `bt__glucose` rather than read from a column - target dataset: `None`; the target is not read from a dataset directly - `benchmark_target_field`: `tyg` - this is the name the Benchmark Track sees - target dataset the formula reads: `blood_tests` - target source: `weizmann` - target unit: unitless index - default research stage: `00_00_visit` - valid range: `6.0` to `13.0` - primary metric: `r2` The component windows are applied before computing the product: triglycerides `20` to `1500` mg/dL and glucose `40` to `400` mg/dL. `bt__glucose` follows the `FpgTask` convention and is treated as fasting glucose. ## HPP data source The triglyceride and glucose inputs load from the pinned Weizmann `blood_tests` source. The live operational source HMO table was audited as a candidate replacement. **Audit decision: keep Weizmann.** Agreement is strong, but operational source + `events` yields net −19.0% baseline coverage with ~36% cohort turnover and no offsetting gain. Rationale and reconsideration triggers: blood-test source audit §9. ## Why This Matters TyG is cheap, scalable, and biologically adjacent to insulin resistance. It is not a clamp measurement and should not be read as a direct insulin-resistance phenotype. In PhenoBench it is useful as a reporting target and as a candidate metabolic axis for probing whether embeddings add signal beyond demographics. The §P CGM-side contribution check found that TyG should not currently count as a clean CGM-sleep bimodal target. CGM's apparent TyG signal is explained by the fasting-glucose factor: on the pre-registered screen, CGM's incremental ΔR² over `demo + ln(FPG)` was **−0.000, 95% CI [−0.032, +0.029]** (n = 353 validation participants), against a raw CGM lift over demographics of +0.040 CI [−0.011, +0.094] which itself crosses zero. Sleep carried independent signal (+0.066, CI [+0.022, +0.110]), so TyG remains reportable but not the mechanism-bimodal slot. Derivation and the frozen pass rule are in the foundation_models sweep cited below, **not in this repository**. **What a row here does not support.** TyG is a formula over two fasting laboratory values. Predicting it well is not a measurement of insulin resistance, is not a clamp-equivalent result, and is not a statement about diabetes risk. Because the formula contains fasting glucose outright, a glucose-adjacent modality can score well by reconstructing one factor - which is exactly what the screen above measured and found. ## Related HPP Measurements Demonstrated in HPP: - **IR surrogates do not identify the same people.** Shilo et al. analysed TyG alongside TyG-BMI, METS-IR, TG/HDL and FPG in 10,114 non-diabetic HPP adults aged 35-75 and found them *moderately inter-correlated but capturing distinct metabolic facets* (Shilo et al. 2026). So a result on TyG does not transfer to "insulin resistance" as a construct, and it does not transfer to another surrogate. - **CGM predicts visceral fat weakly, and anthropometry predicts it well.** In the same cohort, DEXA-derived VAT is predicted at R² = 0.659 from anthropometrics alone but only R² = 0.078 from CGM features, with glycemic *variability* markers outperforming mean-glucose metrics (Shilo et al. 2026). That is the adiposity-dominance this card's floor discussion assumes, measured. - 1.2% of the HPP cohort carry elevated visceral adiposity despite both a normal BMI and a normoglycemic HbA1c (Shilo et al. 2026) - the discordant group any demographics-plus-BMI floor is blind to by construction. **The committed corpus adds a second, larger measurement that agrees with the screen's direction on CGM and disagrees on sleep.** Against the demographic floor: DXA **ΔR2 = +0.182 (95% CI [+0.124, +0.240])** on n = 3,514 clears comfortably, while sleep **+0.016 ([-0.012, +0.047])** on n = 4,036 does not. The foundation_models screen found sleep carrying independent TyG signal on 353 validation participants (+0.066 [+0.022, +0.110]); this larger committed cohort does not reproduce that. Two different cohorts, splits and feature routes, so neither supersedes the other - but the sleep claim is not settled, and the card should not be read as though it were. Proposed rather than shown: that a learned representation adds TyG signal beyond demographics *and* beyond `ln(FPG)`. The screen answers that for CGM only. ## Clinically Meaningful Variants **No categorical TyG variant is defensible, and the reason is stronger here than for a measured analyte: TyG has no diagnostic threshold to transfer.** Published TyG cutoffs for identifying insulin resistance are derived per study population against a local reference standard, and they move with the assay units of both components, the fasting protocol, and the population's adiposity distribution. There is no guideline cutoff comparable to FPG ≥126 mg/dL or HbA1c ≥6.5%. Two further blocks specific to this Task: - **The index inherits two assays.** A threshold applied to a TyG computed from a different triglyceride or glucose method is not the same threshold. The components here are pinned to Weizmann `blood_tests`; a source swap changes the index distribution without changing the formula. - **The construct is not the target.** Since HPP's own surrogates disagree with each other (above), binning TyG would produce an "insulin resistant" label that neither the clamp nor another surrogate would reproduce. If a categorical insulin-resistance Task is wanted, the defensible route is a measured one - fasting insulin or a clamp-anchored reference where available - not a bin on this index. That is the open question below, already recorded. ## Baselines And Ceilings The default floor is a no-modality-feature baseline: `NoFeaturesPredictor` contributes no modality features, and `DemographicFloorStrategy` fits age, sex, and BMI. This matters because adiposity can dominate insulin-resistance surrogates; any embedding result should be interpreted as lift over that demographic floor, not as raw TyG predictability. For CGM-specific interpretation, compare against a model that controls for `ln(FPG)`. A CGM lift that disappears after this control is glucose reconstruction, not evidence that CGM encodes the triglyceride component. ## Caveats - Formula-derived target: TyG includes glucose directly, so CGM predictors can look useful by reconstructing the glucose term. - Fasting assumption: `bt__glucose` is treated as fasting glucose by convention; any upstream source drift would change interpretation. - Insulin-resistance proxy: TyG is not clamp-measured insulin resistance. - Adiposity confounding: demographic and BMI signal should be expected. - Cohort intersections can be much smaller than the full TG/FPG cohort when frozen CGM and sleep embedding caches are required. - In the 2026-06-20 CGM/sleep screen, `n_tyg` is validation/eval n, not the full train+val accessible cohort. ## Metrics Primary metric is `r2` on the configured eval split. For modality-claiming work, report demographic-floor ΔR² and, for CGM, a nested `demo + ln(FPG)` control. For CGM-sleep claims, use paired within-target comparisons rather than cross-target R² subtraction. ## References - Simental-Mendía et al., 2008: original TyG index. - Guerrero-Romero et al., 2010: TyG validation against clamp-derived insulin resistance. - Shilo et al. (2026), heterogeneity of insulin-resistance surrogates in 10,114 non-diabetic HPP adults, *medRxiv*, . - **foundation_models** `sweeps/tyg_cgm_contribution_check_20260620/iteration_summary.md`: the frozen pre-registration and the measured CGM-side contribution screen (C1-C4 criteria, results run 2026-06-20). **That path is in the foundation_models repository, not phenobench** - an earlier version of this card cited it as if it resolved here, and it does not. ## Open Questions - Whether a measured-insulin surrogate should replace TyG as the next mechanism-bimodal candidate. - Whether future cache expansion changes the CGM-over-`ln(FPG)` result enough to reopen TyG's bimodal-target status.