Opens the ir-surrogates cluster. The triglyceride-glucose (TyG) index is a cheap fasting readout of
insulin resistance that PREDICTS cardiovascular events in the general population — but its predictive
signal splits by outcome, and it is a marker to stratify with, not a lever to pull. TyG is computed
from two routine labs: «The TyG index was cal- culated using the following equation: TyG = Ln (TG [mg/
dL] × fasting glucose [mg/dL]/2)»
(Liu et al., 2022).
Liu positions it as a surrogate for insulin resistance itself: the TyG index «is considered a reliable
surrogate marker of insulin resistance», an accessible alternative to the hyperinsulinemic-euglycemic
clamp
(Liu et al., 2022).
So TyG is a surrogate for a surrogate: a lab readout that stands in for insulin resistance, which
is itself a driver several steps upstream of an event.
The decision-relevant finding: incidence-positive, mortality-null type-C (provisional)
Liu’s own headline is a split that a one-line TyG predicts CVD gloss erases. Highest vs lowest TyG category, general population, 12 cohort studies / 6,354,990 participants (no RCT met inclusion):
| Outcome | HR (highest vs lowest) | 95% CI | I2 | studies (n) | GRADE certainty |
|---|---|---|---|---|---|
| Coronary artery disease (CAD) | 2.01 | 1.68–2.40 | 0% | 3 (30,054) | moderate |
| Composite CVD (incidence) | 1.46 | 1.23–1.74 | 82% | 5 (259,757) | very low |
| Myocardial infarction (MI) | 1.36 | 1.18–1.56 | 35% | 2 (5,614,862) | very low |
| Cardiovascular mortality | 1.10 | 0.82–1.47 | 76% | 3 (377,091) | very low |
| All-cause mortality | 1.08 | 0.92–1.27 | 87% | 4 (521,694) | very low |
«Compared with the lowest TyG index category, the highest TyG index was related to a higher incidence of coronary artery disease (CAD) … myocardial infarction (MI) … and composite cardio- vascular disease (CVD) … However, there was no association between the TyG index and mortality (cardiovascular mortality … or all-cause mortality …)» (Liu et al., 2022). The two mortality CIs both cross 1 — this is a no-meaningful-association result on the outcomes people weight most, not merely a weaker positive. Do not round it up to predicts mortality. Whether the null is a true absence or a power/follow-up artifact is open — Liu attributes it to «the limited number of included studies or insufficient follow-up time» and notes other (non-general) populations do show a mortality signal. Shape is therefore outcome-specific, exactly as on Surrogate Outcomes. And the mortality-null is not uniform across IR readouts: the HOMA-IR composite (insulin x glucose) shows a modest mortality signal where both TyG and fasting insulin alone are null-to-borderline — though whether that gap is the measure or the population (Zhang’s non-diabetic Western pool vs the general/Asian TyG pool) is unidentified (Zhang 2017, below).
Effect estimate
- Effect measure (relative only). Categorical HRs above; dose-response per 1-unit TyG increase: CAD HR 1.35 (95% CI 1.02–1.79, I2=94%), CVD HR 1.23 (95% CI 1.12–1.36, I2=89%). Liu reports no absolute risks — a large relative HR on an unstated baseline is not yet a decision; the absolute layer is a G-gap here -> Baseline Risk and the Relative-Absolute Split.
- Population / comparator. Adult general population (age >18), free of established CVD at baseline; highest vs lowest TyG category. NOT a diabetic / established-disease population (those are excluded).
- Outcome. CAD, MI, composite CVD incidence (positive); CV and all-cause mortality (null).
- Dose-response shape. Positive and «linear» over the studied range (P-nonlinear 0.3807 CAD, 0.0612 CVD) (Liu et al., 2022). The studied TyG range spans roughly 6.4–12.4 (category cutpoints in Table 1); the linear read is only «over the studied range» — no threshold or knee is located, and Liu explicitly leaves cutoff values to future work. Monotone-over-range is weak evidence of a true monotone curve (a single per-category display cannot show a knee).
- Uncertainty. Wide, and driven by high heterogeneity (I2 up to 87–94% on the pooled and dose curves) plus all-observational design. GRADE: moderate for CAD only (upgraded for a large effect and dose-response), very low for CVD, MI, and both mortality outcomes: «GRADE assessment indicated very low certainty for CVD, MI, cardiovascular mortality and all-cause mortality, and moderate certainty for CAD» (Liu et al., 2022).
- Effect modifiers (candidate, not established). Diabetes and sex are flagged by cited single studies, not by Liu’s own pooled interaction test — hold as route-(b) hypotheses, not findings (below).
- Certainty: the source’s own GRADE (above).
Ding 2021 — a second gold TyG MA: incidence corroborated, stroke added, shared-cohort (F not E) type-F
Ding 2021 is a gold MA of eight cohort studies / 5,731,294 participants «without ASCVDs at baseline» — the ASCVD-incidence counterpart to Liu’s mortality-inclusive pool. TyG is framed identically, as «the triglyceride–glucose (TyG) index, a novel surrogate indicator of insulin resistance» (Ding et al., 2021). It corroborates and extends the incidence arm but does not independently re-test it (shared constituent cohorts — verdict below), so this is type-F refinement / shared-evidence corroboration, NOT type-E independence. Matched-parameter table (each cell quoted; both pools are general-population, no established CVD at baseline, highest-vs-lowest unless noted):
| Parameter | Ding 2021 | Liu 2022 | Same quantity? |
|---|---|---|---|
| Composite incidence | ASCVD HR 1.61 (1.29–2.01), I2=80%, 5 studies | CVD HR 1.46 (1.23–1.74), I2=82%, 5 studies | concordant estimates, but composites may differ (Ding ASCVD = CAD+stroke+PAD; Liu CVD composite not verified identical) |
| CAD | HR 1.95 (1.47–2.58), I2=92%, 6 studies | HR 2.01 (1.68–2.40), I2=0%, 3 studies | yes — CAD incidence; point estimates concordant (heterogeneity differs) |
| Stroke | HR 1.26 (1.23–1.29), I2=0%, 3 studies | not pooled | Ding-only — fills a Liu G-gap |
| Per-1-unit composite | ASCVD HR 1.28 (1.13–1.45), I2=61%, 4 studies | CVD HR 1.23 (1.12–1.36), I2=89% | concordant slope; same composite-definition caveat as above |
| Per-1-unit CAD | HR 1.39 (1.18–1.64), I2=89%, 3 studies | HR 1.35 (1.02–1.79), I2=94% | yes — per-unit CAD slope, concordant |
| CV / all-cause mortality | not pooled (incidence MA) | 1.10 (0.82–1.47) / 1.08 (0.92–1.27), null | Ding SILENT — no rebuttal of the null |
| Effect modification | pooled subgroup: no sig. modification by age/sex/diabetes (all P>0.05) | cited single-study hints only | Ding refines toward null via its own pooled test |
- The additive contribution — the stroke arm. «the participants with the highest TyG index category had a significantly increased risk of stroke during follow-up compared to those with the lowest TyG category (HR: 1.26, 95% CI 1.23–1.29, I2 = 0%, P < 0.001; Fig. 5)» (Ding et al., 2021). Liu pooled no stroke outcome, so this is a genuine gap-fill — the incidence-positive signal now spans CAD and stroke. Caveat: the near-zero I2 is not three-study agreement — Ding flags the pool is «mainly driven by the result of this study» (one large Korean cohort weighting far above the other two) (Ding et al., 2021), so the tight CI reflects one cohort’s weight, not convergence. Treat the stroke point estimate as effectively single-cohort.
- Per-1-unit dose-response slope, concordant but shape still unclaimed. Composite per-unit HR 1.28 (1.13–1.45) and CAD per-unit HR 1.39 (1.18–1.64) (Ding et al., 2021) track Liu’s per-unit slopes (CVD 1.23, CAD 1.35) — two pools now agree a positive per-unit gradient exists. But Ding is more cautious on shape than Liu: «it remains unknown whether the association between the TyG index and an increased risk of ASCVDs is linear and what the optimal cut-off value of the TyG index is» (Ding et al., 2021). So the monotone-over-range read is not upgraded to a true monotone curve, and no knee/cutpoint is located by either MA. (Abstract/body discrepancy: Ding’s abstract labels 1.39 [1.18–1.64, I2=89%] as the ASCVD-continuous slope; the Results body assigns that to CAD-continuous [3 studies] and gives ASCVD-continuous as 1.28 [1.13–1.45, I2=61%, 4 studies]. Body values used above.)
- CAD and composite point estimates corroborated across both pools — CAD 1.95 vs Liu 2.01; composite 1.61 vs Liu CVD 1.46 (concordant, though the composites may not be the identical construct — Ding’s ASCVD is CAD+stroke+PAD, Liu’s CVD composite is unverified here) — a broader pooling, not an independent replication (see the verdict below).
- Effect modification refined toward null. Where Liu carried only cited single-study hints (diabetes, sex), Ding ran its own pooled subgroup tests: the raised-risk association «were independent of the age, sex, or diabetic status of the participants (for subgroup analyses, all P > 0.05)» (Ding et al., 2021). So the route-(b) hypotheses now run weakly against modification, not merely untested — do not personalize the incidence signal on age/sex/diabetes.
- Silent on mortality; causation still disclaimed; no GRADE. Ding pools only incidence outcomes (CAD, stroke, composite — fatal and non-fatal events), so it leaves Liu’s mortality-null untouched, neither corroborating nor rebutting it. Quality by Newcastle–Ottawa, not GRADE: «The Newcastle–Ottawa Scale score was nine for all of the included studies, indicating good study quality» (Ding et al., 2021), and «this meta-analysis was based on cohort studies; thus, a causative association … cannot be implied» (Ding et al., 2021). Ding leaves risk-score incrementality untested («further studies are needed to determine whether the addition of the TyG index to conventional ASCVD risk prediction tools, such as the Framingham risk score, can improve the predictive efficacy») (Ding et al., 2021) — weaker than Liu’s cited FRS-non-improvement finding, but same direction: neither MA shows demonstrated incremental value over an existing score.
Independence verdict — F, not E [shared-cohort]. Both MAs pool general-population TyG->CVD cohort
studies from PubMed/Embase over overlapping windows (Ding through Jan 2021; Liu later and larger — 12
cohorts / 6.35M vs Ding’s 8 / 5.73M), so Liu is very likely a superset of Ding’s constituents; the
large Korean cohort dominating Ding’s stroke pool is exactly the class Liu’s larger CVD/MI pools draw
on. The included-study lists cannot be shown disjoint, so the two are corroboration on shared
evidence, not independent backing — no [E-independent] token. The incidence arm is therefore more
broadly pooled, not more independently confirmed; confidence: stays low.
(Ding et al., 2021; inferred from Liu et al., 2022)
Zhang 2017 — the direct-insulin arm: the mortality signal is HOMA-IR-composite-specific and fragile type-F
Zhang 2017 is a moderate-tier MA of 7 prospective cohorts / 26,976 non-diabetic adults that pools the direct IR measures — fasting insulin and HOMA-IR (insulin x glucose) — against CV and all-cause mortality, the outcomes on which Liu’s TyG proxy came up null. It does not use TyG, so it approaches the node’s mortality question from a different measure and a (very likely) disjoint study pool. Its headline: «IR as measured by HOMA-IR but not fasting insulin appears to be indepen- dently associated with greater risk of cardiovascular or all-cause mortality in non-diabetic adults» (Zhang et al., 2017).
The node now holds three IR measures on mortality, and they do not agree — only the HOMA-IR composite shows a clear positive, while fasting insulin alone and TyG are null-to-borderline; the signal is fragile and low-certainty throughout, and whether the HOMA-IR-vs-TyG gap reflects the measure or the population is unidentified (they are confounded). Matched-parameter table (each cell quoted; highest-vs-lowest category throughout):
| Parameter | Zhang 2017 (direct: HOMA-IR / fasting insulin) | Liu 2022 (proxy: TyG) | Same quantity? |
|---|---|---|---|
| All-cause mortality | HOMA-IR RR 1.34 (1.11-1.62), P=0.002, 4 studies, I2=44%; fasting insulin RR 1.13 (1.00-1.27), P=0.058, 3 studies, I2=11% | TyG HR 1.08 (0.92-1.27), 4 studies, I2=87% | same OUTCOME; NO on the relation — different EXPOSURE (direct insulin/HOMA-IR vs TG-based proxy) + different POPULATION (non-diabetic-only vs general) |
| CV mortality | HOMA-IR RR 2.11 (1.01-4.41), P=0.048, 2 studies, I2=75%; fasting insulin RR 1.40 (0.49-3.96), P=0.526, 1 study | TyG HR 1.10 (0.82-1.47), 3 studies, I2=76% | same outcome; NO (same exposure/population reasons); both arms underpowered (1-3 studies, wide CIs) |
| Exposure construct | fasting insulin; HOMA-IR = insulin(uU/ml) x glucose /22.5 | TyG = Ln(TG x glucose /2) | NO — both composite-with-glucose, but the non-glucose term differs (insulin vs triglycerides) |
| Study pool | 7 insulin/HOMA-IR cohorts, 2000-2015, mostly Western (3 US / 3 EU / 1 Korea) | 12 TyG cohorts, 2014-2021, 10/12 Asian | disjoint — a TyG study needs TG+glucose, an insulin study needs assayed insulin; different papers |
- The positive is HOMA-IR-composite-specific, NOT a smooth TyG->insulin gradient. Only the HOMA-IR composite (insulin x glucose) shows a clear all-cause-mortality signal: RR 1.34 (P=0.002). Fasting insulin alone is a borderline-null (RR 1.13, P=0.058, CI touches 1) and TyG is null (HR 1.08) — so fasting insulin patterns with the null TyG end, NOT with HOMA-IR, exactly as Zhang’s own «IR as measured by HOMA-IR but not fasting insulin appears to be indepen- dently associated» headline says (Zhang et al., 2017). Zhang’s stated reason is composite-specific: «IR is considered a better biomarker than insulin or glucose alone because it incorporates both the biomarkers» (Zhang et al., 2017). So do not read a gradient toward insulin: the intermediate step (fasting insulin) is statistically indistinguishable from the null end. And even the one positive arm is a 34% relative excess at low certainty; CV-mortality rests on 1-2 studies Zhang itself calls «unreliable due to the small number of articles included» (Zhang et al., 2017).
- Distinction, not a tension
[not-joined (ii)]. TyG-null vs HOMA-IR-positive on mortality is a clash of different quantities in different populations, not a joined issue: different exposure construct, different (non-diabetic-restricted vs general) population, disjoint study pools, different era/follow-up. Once matched, the two are not contradictory — so this is recorded as a non-uniformity distinction (an F-refinement), not a filed[[tension]]. But the same confounds cut both ways: they mean the TyG-null-vs-HOMA-IR-positive gap cannot be attributed to the measure alone — population, pool, era and follow-up are all confounded with the measure, and no study compares readouts head-to-head in one cohort. The identified claim is only that the mortality signal is non-uniform across IR readouts/populations; whether the driver is the measure or the population is unidentified. Zhang reinforces the limit: even within its own pool it cannot compare HOMA-IR to fasting insulin head-to-head — «the result was established on an indirect comparison … most studies failed to report results simultaneously» (Zhang et al., 2017). - Independence verdict — F, more independent than Ding but NOT E
[different-outcome]. Zhang is methodologically more independent than Ding was: a different exposure measure and an almost-certainly disjoint study pool (unlike Ding, which shared constituent cohorts with Liu). But E requires the independent route to reach the same claim, and Zhang lands on a different outcome (mortality, where the measures diverge) than Liu/Ding’s incidence-positive signal — so no single claim is jointly and independently confirmed. No[E-independent]token; the addition is type-F (it refines the node’s mortality picture and bounds the mortality-null as partly proxy-specific).confidence:stayslow: a fragile, non-uniform, low-certainty signal across few studies with wide CIs is not a basis to raise it. - Mechanism corroborated at direction level only. Zhang invokes the same chain Liu does — «IR promotes the development of atherosclerosis through increasing insulin and glucose lev- els. Hyperinsulinemia and hyperglycemia can exert direct atherogenic effect on the vessel wall» (Zhang et al., 2017) plus IR-driven proatherogenic lipids and adipose inflammation — directional support for the readout framing, not an outcome finding.
(inferred from Liu et al., 2022; Zhang et al., 2017)
Prediction is not causation — the framing that keeps this a marker type-C (provisional)
TyG belongs on the predictor side of the Surrogate Outcomes prognostic-marker-vs-lever line. Two of Liu’s own statements pin it there:
- It adds nothing over an existing risk score. «On the other hand, addition of the TyG index to the Framingham Risk Score (FRS) did not lead to improvement in its predictive power; thus, adding the TyG index to the FRS does not improve CVD risk predic- tion» (Liu et al., 2022). A predictor that does not improve an incumbent’s discrimination is not carrying independent decision value once the standard factors are in hand.
- Causation is disclaimed. All 12 studies are observational cohorts: «residual confounding factors may have influenced our results; thus, causation cannot be proven» (Liu et al., 2022). Liu’s strongest positive claim is that TyG «may be considered an independent predictor for CVD incidence» (Liu et al., 2022) — predictor, never target.
The synthesis move (this page’s own): a raised TyG is a readout of the atherogenic-dyslipidemia + insulin-resistance state, not a rival causal lever competing with LDL/apoB. Liu’s mechanism section says so — «Insulin resistance in liver and adipose tissues drives the development of atheroscle- rotic dyslipidemia, generates a low-grade inflammatory state, and increases release of inflammatory markers» (Liu et al., 2022). That atherogenic dyslipidemia is apoB-bearing particles (small-dense LDL, VLDL remnants) — the very burden the causal-lipid nucleus tracks. So TyG and the causal apoB story are not competitors: TyG flags the discordant stratum where LDL-C under-states the apoB particle count, and apoB remains the causal quantity and the target -> LDL ApoB and Cumulative Exposure. Reading TyG as insulin/triglycerides beat LDL confuses a downstream marker of the atherogenic state for the causal driver within it. (inferred from Liu et al., 2022)
Mechanism (marked, directional)
Insulin resistance -> low-grade systemic inflammation -> endothelial dysfunction, plus IR-driven atherogenic dyslipidemia and plaque progression; TyG is a lab proxy for the IR at the head of that chain. Liu grounds direction in a genetic study (63,746 CAD cases / 130,681 controls) implicating lipid metabolism and inflammation in coronary atherogenesis, and in TyG’s strong concordance with the clamp (reported sensitivity 96.5%, specificity 85.0%) and its edge over HOMA-IR — these figures are the source’s (Liu et al., 2022). The directional reading below is the wiki’s. This is mechanism with human corroboration — admitted directionally, not as an outcome finding; whole-organism net effect (not the intended mechanism) still governs.
Decision relevance
- Use TyG to place, not to treat. It is a cheap route-(a) stratifier — genuinely useful for reading who sits in the insulin-resistant / atherogenic-dyslipidemia stratum from two routine labs — but it is not a validated target, and it does not improve an existing risk score’s discrimination. Steering someone by lower your TyG has no intervention evidence behind it here.
- A raised TyG is a prompt to check apoB, not a rival to LDL lowering. In the insulin-resistant, hypertriglyceridemic person, LDL-C can understate the apoB particle burden; TyG flags exactly that discordant stratum -> LDL ApoB and Cumulative Exposure. The lever remains apoB-lowering (and the upstream ectopic-fat / energy-balance levers), not the marker.
- Weight the mortality null honestly; it is not uniform across IR readouts. For TyG the pooled mortality signal is absent (both CIs cross 1) at very-low certainty — no mortality claim can be built on TyG from this evidence. But the null is not flat across readouts: the HOMA-IR composite carries a modest all-cause signal (RR 1.34), while fasting insulin alone is only borderline (RR 1.13, P=0.058) and patterns with the null TyG — so the positive is specific to the HOMA-IR composite, not to IR readouts generally (Zhang 2017, moderate-tier, low-certainty). Whether even that gap is the measure or the non-diabetic-vs-general population is unidentified. The honest composite is a fragile, non-uniform mortality signal — neither a flat null nor a clean gradient — still not a target, and still weaker than the incidence signal and than the apoB causal evidence.
(inferred from Liu et al., 2022; Zhang et al., 2017) — the decisions in this section (use-to-place-not-treat, prompt-to-check-apoB, weight-the-null) are the wiki’s application of Liu’s effect estimates and FRS/causation findings, and of Zhang’s direct-measure mortality estimates, quoted above — not additional source claims.
Limits
- All observational, causation disclaimed; no RCT. 12 cohorts, mostly Asian (10/12), 2014–2021; moderate certainty on CAD alone, very low on the other four outcomes. (Study characteristics and heterogeneity below are the source’s (Liu et al., 2022); the R1 coherence note is the wiki’s.)
- High heterogeneity on CVD and both mortality pools (I2 76–94%); subgroup and publication-bias analyses were not run (fewer than 10 studies per outcome).
- Diabetes and sex as route-(b) candidates only — now weakly against. Liu’s cited single studies suggest TyG-CAD holds in non-diabetics but attenuates in diabetics (a route-(b) effect-modification hint), and a sex difference in subclinical atherosclerosis — but Liu’s own pooled interaction tests were underpowered. Ding 2021 then ran its own pooled subgroup tests and found the incidence association independent of age, sex, and diabetic status (all P>0.05, above), so the route-(b) hints now run weakly against modification. Do not personalize the incidence signal on them.
- Coherence, not validity (R1): this node is internally coherent and source-faithful; no operation here grades it against a realized outcome. The loop is open.
CASHED 2026-08-09: Ding 2021 ingested (see the Ding 2021 — a second gold TyG MA section above). The included-studies overlap check resolved the way the hold anticipated — shared constituent cohorts are likely (both general-population PubMed/Embase TyG pools over overlapping windows; Liu later and larger), so Ding corroborates the incidence-positive arm as type-F (shared evidence), not E-independent, and adds the stroke outcome Liu lacked.
CASHED 2026-08-09 (partial): Zhang 2017 ingested (see the Zhang 2017 — the direct-insulin arm section above) — the direct measures fasting insulin / HOMA-IR against mortality. It partially answers the TyG-specific-vs-general split from the direct-measure side: the mortality-null is not uniform across readouts, since the HOMA-IR composite retains a modest mortality signal (RR 1.34) where fasting insulin alone and TyG are null-to-borderline — HOMA-IR-composite-specific and fragile, not flat across IR readouts (measure-vs-population attribution unidentified).
AWAITS a TG/HDL-ratio outcome source — to lift the cluster from two IR readouts (TyG, direct insulin/HOMA-IR) to the neutral question across surrogates, and to test the incidence/mortality split on the TG/HDL ratio arm specifically (Zhang covered the direct-insulin arm, not TG/HDL).