When Does Visceral Fat Change the Right Treatment?

Doctors increasingly measure body composition in detail rather than relying on body weight. They can now split a body into visceral fat (the fat packed around the organs), subcutaneous fat (the fat under the skin) and skeletal muscle, often for free, from CT scans that were taken for some other reason. The literature on this is almost entirely prognostic. It asks whether a high visceral-fat ratio predicts relapse or death, and it answers with an area under the ROC curve. A treatment decision needs a different question answered: would knowing this patient's visceral fat or muscle mass change which drug we should give now, or what we should do next if the drug fails? A variable can predict outcomes well and still never change the best choice. That happens when it makes every option look equally bad.

We propose to measure and test that second quantity, the decision value of granular body composition, inside dynamic treatment regimes. A dynamic treatment regime is a sequence of rules that map a patient's history so far to the next treatment. The core claim is that body composition earns a place in a regime only as a tailoring variable: something that modifies how much a treatment helps, enough to flip which treatment is best for a meaningful fraction of patients. There is already a hint that this is the right lens. In a Crohn's disease cohort, baseline visceral fat ratios predicted one-year relapse under 5-aminosalicylic acid (AUC 0.849) and under steroids plus azathioprine (AUC 0.820), but showed no association at all under infliximab (Can visceral adipose tissue and skeletal muscle predict recurrence…). That is a treatment-by-composition interaction hiding inside a paper that reported it as three separate prognostic results. We will build the estimation and evaluation machinery to find such interactions, check that they hold up, and say when they matter, under the conditions where body composition is actually available: measured rarely, irregularly, and with error.

Core Claim and What Is New

Write the stage-$t$ quality function as $Q_t(h_t, a)$: the expected long-run outcome given history $h_t$ and binary choice $a \in \{0,1\}$ (for example conventional therapy versus biologic, or hold versus escalate). The contrast is

$$ C_t(h_t) = Q_t(h_t,1) - Q_t(h_t,0) = \psi_0 + \psi_{\text{BMI}}\,\text{BMI}_t + \psi_{V}\,\widehat{\text{VAT}}_t + \psi_{M}\,\widehat{\text{SMI}}_t + \dots $$

Here $\widehat{\text{VAT}}_t$ and $\widehat{\text{SMI}}_t$ are the current best estimates of visceral fat and skeletal muscle index, carried forward from the most recent scan together with their uncertainty. The optimal rule treats when $C_t > 0$. The decision value of body composition is the gain in mean outcome from using a rule that can see composition, compared with the best rule that can only see BMI and routine labs.

This gain is nonzero only under two conditions. The interaction coefficients $\psi_V$ or $\psi_M$ must be nonzero, and the sign of $C_t$ must actually flip within the range of composition seen in real patients. A purely prognostic effect of visceral fat shifts $Q_t(h,0)$ and $Q_t(h,1)$ equally, so it adds nothing. This is why the AUCs in the current literature cannot tell us whether segmentation pipelines belong in treatment protocols.

Three things are new relative to the file's closest work. The Crohn's studies (recurrence; Can visceral adipose tissue and skeletal muscle predict recurrence of newly diagnosed Crohn’s disease in different treatments; infliximab radiomics) use a single pre-treatment scan to predict a fixed-therapy outcome. They never estimate a decision rule, and they never use follow-up composition. The tacrolimus dosing model (AI-Driven Tacrolimus Dosing) and the ventilation weaning policy (A Reinforcement Learning Approach to Weaning…) are sequential, but they use only aggregate weight. We treat composition as a slow, sparsely observed part of the state. We estimate its interaction with treatment using low-dimensional, inference-ready Q-learning rather than deep offline reinforcement learning. We evaluate the result by decision value, not by prediction accuracy.

Research Program

Measurement layer. The first component turns imaging into a time-indexed composition state. We run Comp2Comp on every abdominal CT already in the record. In internal validation it reports Dice scores of 0.97 for muscle, 0.96 for visceral fat and 0.97 for subcutaneous fat at L3, with area errors averaging below 2%. Comp2Comp gives us visceral fat area, subcutaneous fat area, muscle area, intermuscular fat and their ratios.

Between scans, composition is carried forward by a simple state-space model. Its noise variance grows with time since the last scan, and that variance is calibrated against patients who had repeat scans. We will use the Comp2Comp external-validation Dice (above 0.80 to 0.85) as a lower bound on measurement quality off-site.

A surface-geometry proxy in the style of AbdCTBench is treated as a second, noisier channel. Its known weaknesses are deliberately carried into the error model rather than assumed away: it is single-site, has no validation on consumer depth sensors, and its error is larger for women. This matters because repeated CT is not ethical for monitoring, so any realistic regime will run on noisy proxies.

Estimation layer. The second component estimates the contrast functions by backward Q-learning with variable selection, in the style of sequential advantage selection. We use the asymptotic inference results for multi-stage stationary policies with variable selection to obtain confidence intervals on $\psi_V$ and $\psi_M$.

We choose a parsimonious estimator deliberately. EpiCare shows that offline deep reinforcement learning needs more than about 2,000 episodes to beat a standard-of-care heuristic, and falls below random under 256. The real cohorts in this area have 171 to 231 patients.

Follow-up composition is both a consequence of earlier treatment and a confounder of later treatment. Steroids waste muscle, and biologics may shrink mesenteric fat. To avoid post-treatment bias we adjust for it using the transition-kernel, Neyman-orthogonal approach of Modeling Covariate Transition…. Where we need counterfactual trajectories under a candidate rule, we forward-simulate with G-Net.

Evaluation layer. The third component estimates decision value without trusting standard off-policy evaluation, which the file shows is unreliable. In EpiCare the Direct Method had a root-mean-square error of 106.4 on a return bounded by 100. In the sepsis stress tests (Position: RL in DTRs Needs Critical Reexamination), random policies sometimes "won" under standard metrics.

Our main evidence therefore comes from simulators with known ground truth. In real data we report only estimands that need weaker assumptions: the interaction coefficients with confidence intervals, and the fraction of patients whose recommended action changes when composition is added. We report policy values only with doubly robust intervals and explicit overlap diagnostics.

Baselines

The naive reference point is the clinician or standard-of-care rule. In the simulators this is the heuristic policies shipped with EpiCare and DTR-Bench. In real data it is the observed treatment sequence. Any regime that cannot beat it on ground-truth return in simulation is discarded.

The BMI-tailored regime is the comparison that defines decision value. It uses the same Q-learning estimator, the same history, the same sample-size and tuning budget, and the same cross-validation folds. The only difference is that it sees BMI, weight and routine labs (albumin included, given its role in the Crohn's cohort) in place of the composition terms. The composition regime beats it if the ground-truth value gain is positive with a 95% interval excluding zero in simulation, and if in real data $\psi_V$ or $\psi_M$ is significant and changes the recommended action for a nontrivial fraction of patients.

The anthropometric regime replaces imaging with tape-measure indices: the weight-adjusted waist index (Yahong Gu et al.), the body roundness index (Zechao He et al.) and the metabolic score for visceral fat (Shiping Li et al.). These are the cheapest granular proxies, so CT-based composition has to beat them in decision value, not merely match them.

The strongest existing static method is baseline-only stratification as in the Crohn's recurrence study. Its reported per-arm prognostic AUCs are 0.849 (5-ASA) and 0.820 (steroids plus azathioprine). The radiomic infliximab-resistance model (CT radiomics of VAT and bowel) reports an external-validation AUC of 0.833. We refit both as single-stage rules on our data. Our regime beats them only if adding follow-up composition and a second decision stage raises value beyond what the best baseline-only rule achieves.

For dosing, the reference is the LSTM in AI-Driven Tacrolimus Dosing, which reports a next-day trough mean absolute error of 1.880 ng/mL overall and 1.521 ng/mL inside the therapeutic window. We retrain it on our cohort with and without composition features before comparing.

The deep offline reinforcement learning competitor is conservative Q-learning (CQL), the best performer in EpiCare. It gets composition in its state and the same tuning budget, alongside behavior cloning.

The key ablation is a prognostic-only model. Here composition enters the outcome model but is barred from the treatment contrast. If this does as well as the full model, composition carries prognostic value but no decision value.

Why not just…?

Why not just measure composition once at baseline and assign a fixed therapy? This is the most serious threat. The Crohn's data suggest a single scan may already identify who should go straight to a biologic, and if body composition changes slowly, follow-up scans may add little. We do not argue around this. The comparison between the static rule and our two-stage regime is a primary outcome, and a null result is reportable: it would say that composition is a baseline triage variable, not a monitoring variable. Static rules do have one known structural weakness. They cannot use mid-treatment changes such as steroid-induced muscle loss without either post-treatment bias or wasted information (Modeling Covariate Transition…). The simulator measures how large such changes must be before the second stage pays off.

Why not just add visceral fat to a supervised outcome model or a behavior-cloning policy? A supervised outcome model gives prognosis, which we argue is the wrong quantity. Behavior cloning copies clinicians, who currently ignore composition, so it cannot discover a composition-based tailoring rule. Behavior cloning is nonetheless a strong policy baseline, since it often matched or beat offline reinforcement learning in the sepsis stress tests, so it is included. The prognostic-only ablation directly tests the claim that prediction gains do not translate into decision gains.

Why not just use tape-measure indices? If the waist-adjusted indices recover most of the decision value, that is the more useful finding clinically, and we would report it as such. The evidence against them is indirect. The Crohn's relapse signal was a visceral-to-total-fat ratio that circumference cannot separate from subcutaneous fat, and the infliximab signal lived in internal wavelet texture. The anthropometric regime baseline settles the question empirically.

Why not just use deep offline reinforcement learning with a richer state? At cohort sizes of a few hundred, EpiCare shows such agents fall below random, and DTR-Bench shows recurrent variants do not reliably help. We include CQL so that its predicted failure at small sample sizes, and any success at large ones, are measured rather than assumed.

Proposed Experiments

Simulation with known ground truth. We modify DTR-Bench's SimGlucoseEnv and EpiCare's liver-disease environment so that a chosen share of each simulated patient's drug-response variability is driven by a composition vector. The vector is sampled from a joint distribution of visceral fat, subcutaneous fat, muscle and BMI fitted to Comp2Comp-segmented CTs. The rest of the variability remains unexplained, as in DTR-Bench's existing perturbations. Under that existing variability, SimGlucoseEnv is reported to give no algorithm a positive reward (DTR-Bench).

We vary four things factorially: the share of variance explained beyond BMI (from zero upward), measurement noise (CT-level versus surface-proxy-level, including the larger error for women), the interval between scans, and the number of training episodes (256, 2,000 and EpiCare's larger datasets). The experiment is circular by design in one respect: if we build composition into the dynamics, composition must help with perfect data. What it does settle is how fast that help erodes under realistic noise, sparse measurement and small samples. It also tests the central conceptual prediction: when composition is made purely prognostic, its predictive AUC can be high while decision value is zero.

Crohn's two-stage regime. We assemble a retrospective cohort shaped like the two published Crohn's cohorts: newly diagnosed patients with a baseline scan (CT or CT enterography) and, where it exists, a follow-up scan at reassessment. Stage one is conventional therapy versus biologic. Stage two is escalate versus continue at nonresponse. The first test is whether the "biologics show no association" finding is a qualitative interaction, meaning the sign of $C_1$ flips across the observed visceral-fat range. The second test is whether follow-up muscle and fat changes enter the stage-two contrast.

Tacrolimus dosing. In a transplant cohort with pre-transplant abdominal CT, we add Comp2Comp composition to the tacrolimus LSTM and to a Q-learning dosing rule. Tacrolimus is fat-soluble, so its spread through the body could plausibly depend on fat and muscle. Two outcomes are tracked: trough prediction error against the 1.880 ng/mL reference, and time in the therapeutic window under the learned rule. In simulation, the decision value is the ground-truth value gap. In real data, we report interaction estimates, action-change rates and doubly robust values with overlap diagnostics.

Decisive checks. Three checks determine the verdict. The proposal's thesis fails if, at realistic noise and sample sizes in simulation, the composition-tailored rule cannot beat the BMI-tailored rule. It also fails if, in the Crohn's data, the confidence interval on the visceral-fat-by-biologic interaction includes zero. It is weakened to a triage claim if baseline-only rules match the two-stage regime. It succeeds if decision value survives proxy-level noise, because only then can surface or opportunistic measurement realistically sit inside a monitoring loop.

References

  1. Can visceral adipose tissue and skeletal muscle predict recurrence of newly diagnosed Crohn’s disease in different treatments
  2. CT-based radiomics signature of visceral adipose tissue and bowel lesions for identifying patients with Crohn’s disease resistant to infliximab
  3. AI-Driven Tacrolimus Dosing in Transplant Care: Cohort Study
  4. A Reinforcement Learning Approach to Weaning of Mechanical Ventilation in Intensive Care Units
  5. Comp2Comp: Open-Source Body Composition Assessment on Computed Tomography
  6. AbdCTBench: Learning Clinical Biomarker Representations from Abdominal Surface Geometry
  7. Sequential advantage selection for optimal treatment regime
  8. Asymptotic Inference for Multi-Stage Stationary Treatment Policy with Variable Selection
  9. EpiCare: A Reinforcement Learning Benchmark for Dynamic Treatment Regimes
  10. Modeling Covariate Transition for Efficient Estimation of Longitudinal Treatment Effects in Randomized Experiments
  11. G-Net: a Recurrent Network Approach to G-Computation for Counterfactual Prediction Under a Dynamic Treatment Regime
  12. Position: Reinforcement Learning in Dynamic Treatment Regimes Needs Critical Reexamination
  13. DTR-Bench: An in silico Environment and Benchmark Platform for Reinforcement Learning Based Dynamic Treatment Regime
  14. Association between cognitive impairment and comprehensive weight phenotypes among older adults in China: a nationwide survey
  15. Association between the body roundness index and all-cause mortality in patients with metabolic dysfunction- associated steatotic liver disease
  16. The metabolic score for visceral fat and risk of peripheral arterial disease in hypertension patients: a prospective cohort study