Do Insulin-Resistance Subgroups Exist? A Survival-Space Null Test for Metabolic Phenotypes

Cheap blood-test formulas for insulin resistance give conflicting answers about who is at risk of dying. The triglyceride–glucose index (TyG) combines fasting triglycerides and fasting glucose. Related scores fold in body mass, cholesterol or waist size. Across studies, the same kind of score predicts death in a U-shape among US adults with diabetes, rises steadily with risk in heart-failure patients, and reverses direction in centenarians. A tempting story is that each cohort is a mix of hidden metabolic subgroups whose risk curves cross over time. The standard way to test that story is to cluster the patients and compare their survival. That procedure is known to invent subgroups where only a smooth gradient exists.

We propose a test that asks whether survival-relevant metabolic subgroups exist at all. It resamples a model that learns subgroups from the outcome, a mixture of Cox survival regressions. It measures how consistently patients land in the same subgroup across resamples. It then compares that consistency against a null world simulated from the best single-population survival model, one with no subgroups but full freedom to bend risk curves smoothly. If real data produce more stable subgroups than any smooth single-population world can, the subgroups are worth naming. If not, the conflicting risk curves are a continuous phenomenon, and the field's growing habit of "metabolic subtyping" should stop. Either outcome is informative.

Core Claim

The obvious study at this intersection transplants the oncology recipe to metabolic markers. That recipe runs consensus clustering (repeatedly clustering random subsamples and recording how often each pair of patients ends up together) on a vector of TyG, the metabolic score for insulin resistance (METS-IR) and the estimated glucose disposal rate (eGDR), then compares Kaplan–Meier curves between clusters. It is what tsRNA-defined subtypes in gastric cancer, CSC-gene subtypes in colorectal cancer and pulmonary LCNEC characterization do on omics data.

We think this recipe is flawed for metabolic markers in two ways.

First, the surrogate indices are deterministic functions of about seven overlapping lab and body measurements. Clusters in index space mostly reflect the shape of the formulas.

Second, a cluster can be geometrically stable and still carry no survival information beyond what a smooth risk curve already captures. M3C fixed the analogous problem in feature space. It simulates correlated Gaussian data with no clusters and shows that naive consensus clustering claimed 7 lupus subtypes where the true count was one. Nobody has built the equivalent null in outcome space.

Our claim is that the correct null for "survival subgroups exist" is a flexible single-population survival model. Stability should be measured on the latent group assignments of an outcome-aware mixture model, not on raw feature clusters.

Research Program

The mixture model is Deep Cox Mixtures (DCM). DCM assumes the population is a mixture of $K$ Cox models. A neural network assigns each patient to a group, and each group gets its own relative risks and its own nonparametric baseline hazard (the underlying event rate over time). DCM is fit by hard-assignment expectation–maximization and reports no stability or null test. We reuse its model class unchanged.

Measuring assignment stability. We fit DCM on $B$ bootstrap resamples at each candidate $K$. Each fit's assignments are converted into a patient-by-patient co-assignment matrix, following the association-matrix lifting of Statistical Guarantees for Consensus Clustering. That paper proves aggregation of such matrices is consistent, and that local refinement drives misclassification down exponentially. Lifting also removes the label-switching problem across runs. Averaging the matrices gives a consensus matrix $M$, with entries

$$M_{ij} = \frac{\#\{b: i,j \text{ both resampled and co-assigned}\}}{\#\{b: i,j \text{ both resampled}\}}.$$

We summarize stability with the proportion of ambiguous clustering, $\mathrm{PAC}_K$. This is the fraction of off-diagonal entries of $M$ that fall between 0.1 and 0.9, as used in the Monti consensus loop.

Building the survival-space null. On real data we fit a single-population Cox model $\hat h_0(t)\exp(g(x,t))$, where $x$ is the patient's covariate vector and $t$ is time since baseline. The function $g$ is deliberately rich: restricted cubic splines in each raw component, spline interactions with age, and coefficients that vary smoothly with time. This is exactly the model that could produce a U-shape or an age-dependent flip without any subgroups. We also fit a censoring model.

We then simulate $R$ datasets that keep the real covariates, draw event times from the fitted survival model, and draw censoring times from the censoring model. The entire bootstrap-DCM-consensus pipeline is rerun on each simulated dataset. The resulting null distribution of $\mathrm{PAC}_K$ gives an empirical $p$-value and an M3C-style relative stability score for each $K$. We choose $K$ as the value with the strongest excess stability over the null, not the lowest raw PAC. The final partition comes from Zhou–Dudeja local refinement on $M$.

Features. The model sees the raw components: triglycerides, fasting glucose, HDL cholesterol, HbA1c, BMI, waist circumference and hypertension status, plus age and sex. The composite indices are recorded only as derived quantities for interpretation.

Computation. For $N\approx 9{,}000$, $M$ has about $8\times 10^7$ entries. That fits in memory, so no sketching is needed at this scale. The cost is the $B\times R$ DCM refits. We plan $B=100$ and $R=200$ small multilayer-perceptron DCMs, which is modest on one GPU.

Baselines

Naive reference. A single standard Cox model with linear terms for TyG, METS-IR and eGDR plus the usual confounders. This mirrors the tertile and continuous analyses of Zhao et al. and Zhou et al.. It fixes the floor on discrimination and calibration.

Strongest competitor for prediction. The flexible single-population Cox model described above. Because it is also our null generator, any claim of subgroups must beat it on held-out data, not merely on stability.

Strongest existing subgroup method. DCM with $K$ picked by validation likelihood and no stability test. On FLCHAIN, DCM reported an expected calibration error of 0.0200 against 0.0992 for standard Cox (Deep Cox Mixtures). Within NHANES metabolic cohorts, no reported result was found on Lacuna for C-index or calibration of any of these models. Zhao et al. report only hazard ratios and spline thresholds (8.95 and 9.00). The closest discrimination figure is the heart-failure cohort's C-statistic rising from 0.690 to 0.729 when METS-IR is added to MAGGIC plus NT-proBNP (Zhou et al.). That cohort is not public, so we cite the figure only for context.

Transplanted oncology pipeline. M3C run on the index vector, followed by Cox regression on the resulting cluster labels.

Ablations. Two ablations isolate our key ingredient. The first replaces the survival-space null with M3C's feature-space null while keeping the DCM stability measurement. The second runs our full pipeline on index-only features instead of raw components.

Fair comparison. Every method uses the same cohort, the same confounder set and the same five-fold outer split. Every tunable method gets the same hyperparameter search budget of 50 configurations, tuned on inner folds.

What counts as beating a baseline. "Subgroups exist" requires two things. The stability $p$-value against the survival null must be below 0.05 after correcting across $K$. The chosen mixture must also improve held-out integrated Brier score or calibration error over the flexible Cox model, with a bootstrap confidence interval excluding zero.

Why not just…?

Why not just fit Cox with restricted cubic splines and age interactions? This is the real threat, and we treat it as the default hypothesis rather than a strawman. Elgui et al. prove that Cox partial likelihood fails to reach the best possible ranking of patients when hazards are not proportional. DCM shows that a single Cox model miscalibrates when latent groups exist. Neither result shows that metabolic cohorts actually contain such groups. Our null is exactly this spline model, so the experiment settles the question directly. If the null reproduces the observed stability, the spline model wins and we report that.

Why not just run DCM and pick $K$ by validation likelihood? Likelihood almost always improves with more components. Hard-assignment expectation–maximization can also split a smooth risk gradient into stable-looking groups. This is the survival-space analogue of the 10-versus-4 glioblastoma overcount in M3C. We will show this directly by running likelihood-selected DCM on the simulated null datasets and counting how often it chooses $K>1$.

Why not just cluster the indices with M3C and compare survival? Feature-space stability answers the wrong question. The indices share inputs; for example, triglycerides and glucose enter TyG, METS-IR and TyG-BMI. Their joint distribution can therefore have lumps created by the formulas that are unrelated to mortality. The index-feature ablation and the survival-irrelevance check below measure this. We also report cases where M3C calls $K>1$ but the clusters add nothing beyond the flexible Cox model.

Proposed Experiments

Primary data. NHANES 1999–2018 with the public linked mortality files. We use two cohorts: the adults with type 2 diabetes studied by Zhao et al. (N=2,998; 883 deaths, 265 cardiovascular), and the larger diabetes-plus-prediabetes population of Wang et al. (N=9,017). Wang et al. analyzed this group cross-sectionally with logistic regression; we add the mortality follow-up. Outcomes are all-cause and cardiovascular death.

Method validation. Before touching NHANES, we check the pipeline on FLCHAIN and SUPPORT, where DCM documented crossing survival curves. We should recover $K>1$ on SUPPORT and approach the reported FLCHAIN calibration gain.

Power check. We run a planted-mixture simulation matched to NHANES sample size and event counts. It contains two groups with crossing hazards of varying separation. This measures what effect size our test can detect, so a negative result can be read as "no subgroups larger than X" rather than "underpowered."

Decisive checks. The first is the survival-null stability test at each $K$. The second is held-out calibration and discrimination against the flexible Cox model, reported as integrated Brier score, calibration error computed as in DCM, and time-dependent C-index. The third tests whether the subgroups explain the paradoxes: whether TyG–mortality curves are monotone within each discovered group even though the pooled curve is U-shaped. The fourth is temporal transport: fit on the 1999–2008 cycles and test whether the group structure and group-specific hazards reproduce on 2009–2018.

Probing the flip. For the age-dependent reversal reported in centenarians (Yan et al.), NHANES top-codes age, so we can only examine whether a separate low-reserve group appears among participants aged 75 and over. We state plainly that NHANES cannot test the centenarian finding itself.