Body Mass Index (BMI) cannot distinguish a person whose weight sits safely under the skin from a person of identical weight whose fat surrounds the internal organs. Visceral fat, the kind packed around the organs, is the version that drives diabetes, liver disease and cardiovascular harm. Measuring it directly requires DXA (a low-dose X-ray body scan), CT or MRI. Cheaper proxies now exist: tape-measure indices such as waist-adjusted-for-weight, 2D body silhouettes, and 3D optical surface scans. All of them are used the same way. One input goes in, one estimate comes out, and nothing decides whether the estimate is trustworthy enough to stop or ambiguous enough to justify the next, more expensive test.
This proposal treats body-composition assessment as a short sequential decision problem. A planner holds a probability distribution over a patient's hidden fat profile. It imagines a small, sampled set of possible test results, a "sparse belief tree" in the style of classical sampling-based planners, and chooses the next test, or stops, to minimize misclassification cost plus monetary and radiation cost. The claim is narrow. A cheap scan's value in a screening cascade is mostly option value: its result decides whether to order the expensive scan. Only a planner that reasons about acting differently after each possible result can represent that value. Myopic one-step rules miss it, and so does selecting a fixed bundle of tests. We test this on cohorts where every modality was measured on every person. Policies can therefore be scored by replaying real measurements, not simulated ones.
Core Claim
Consider a cascade of anthropometrics, then a silhouette or optical-scan estimate, then DXA, then MRI. A closed-loop lookahead planner should reach a given accuracy for classifying visceral adiposity at measurably lower cost than three alternatives: greedy value-of-information rules, re-planned subset selection, and fixed clinical thresholds. The gap should be largest exactly where a cheap test is moderately informative and a costly test is decisive.
The intersection is open. The file found no prior work applying POMDP planning to body composition. A POMDP, or partially observable Markov decision process, is a decision problem where the true state is hidden and must be inferred from observations. Two adjacent methods each assume away the property we target. Cohort-Based Active Modality Acquisition ranks patients for a single pre-chosen expensive modality by expected KL shift in predictions, a one-step rule. Acquisition Conditioned Oracle for Nongreedy Active Feature Acquisition escapes greedy "dead zones" by scoring whole subsets of features with a k-nearest-neighbour conditional model. That subset score assumes the subset is bought regardless of what the first test shows.
The prediction is also refutable. If ACO re-planned after every result matches the tree planner across the cost range, the conclusion is that lookahead is unnecessary here. We would report that outcome as a finding.
Research Program
State and target. The hidden state is a patient's depot profile: visceral fat (VAT), abdominal subcutaneous fat (ASAT), and their ratio. Ground truth is MRI-derived depot volume. Decision targets are binary: "VAT above a sex-specific threshold," and, separately, "VAT/ASAT ratio above threshold." The ratio is the harder target. In Silhouette images enable estimation of body fat distribution and associated cardiometabolic risk, silhouettes predicted VAT well (R² 0.885) but the ratio only moderately (R² 0.553 in males, 0.504 in females). The ratio therefore gives cheap tests more room to be ambiguous. That same paper also motivates the target: the silhouette-predicted ratio carried diabetes risk independent of BMI and waist, with odds ratios of 1.70 in males and 1.74 in females.
Tests. The tests are free anthropometrics (BMI, waist, the weight-adjusted waist index used in Gu et al.), then a silhouette model, then DXA, then MRI. MRI acts as the terminal reveal. Each test has a monetary or time cost. DXA also carries a radiation charge that counts against a hard per-encounter budget.
Observation model. Each "observation" is the scalar output of an existing estimator for that test, not raw pixels. This keeps the generative model low-dimensional and honest. The model is a conditional density, either a mixture density network or a conditional normalizing flow, over not-yet-seen test outputs and the hidden state, given what has been observed so far. It is conditioned on sex. That choice follows Tian et al., who found that learned shape features helped males but not consistently females. It is the place where a sex-blind planner would sample misleading futures for women, so we check calibration per sex.
Planner. The planner represents belief as particles. It expands a tree over the remaining tests, sampling K future observation outcomes per branch. Sampled observations come from fixed per-patient random "scenarios," as in DESPOT. Policy-size regularization guards against overfitting to the K samples. The underlying reason a fixed number of samples per branch gives near-optimal action choices, independent of how fine-grained the state space is, is the sparse-sampling result of Kearns, Mansour and Ng.
Because the horizon is short, sparse sampling here does not buy scale. It buys tractability over continuous test outcomes, plus millisecond-level per-patient decisions. We say this plainly rather than claim large-scale novelty. The radiation budget enters as a hard constraint on admissible branches, not as a soft penalty.
Evaluation by replay. In the evaluation cohorts every patient has every measurement. A policy run on held-out patients reveals each chosen test's actual value. Model mismatch, the main open weakness of sparse planners noted in the file, can then hurt planning quality but cannot contaminate the score.
Baselines
The do-nothing reference classifies from free anthropometrics alone, using a calibrated logistic model with BMI, waist and the weight-adjusted waist index. It costs nothing. For VAT volume, the silhouette paper reports BMI-based linear models at R² 0.608 to 0.833. This is the floor any acquisition policy must clear at non-zero cost.
The acquire-everything reference orders silhouette, DXA and MRI for everyone. It fixes the accuracy ceiling and the maximum cost, and every policy is placed on the cost–accuracy plane between these two points.
The fixed clinical rule orders the silhouette for everyone, then DXA only if predicted VAT falls within a band around the threshold. The band width is tuned on validation data. This is the hand-designed cascade a clinician would propose, and it is a serious competitor because it already encodes "scan if ambiguous."
The greedy value-of-information baseline applies the expected-KL acquisition score of CAMA one step at a time and stops when the score per unit cost falls below a tuned threshold. It is run on the identical conditional density model. No reported AUROC for CAMA was found on Lacuna, and no reported result was found on Lacuna for any method on a body-composition acquisition cascade.
The strongest existing method is ACO. At each step it scores candidate test subsets by expected loss plus cost and acquires the cheapest test in the best subset, then re-plans after seeing the result. We run it twice: with its own k-NN conditional model, and with our density model so that the comparison isolates the decision rule. Its reported strength is matching a full-information recommender "over 90% of the time while only requiring 20 out of 65 possible features" on the Open Bandit Dataset. No reported result was found on Lacuna for ACO on clinical imaging cascades.
Two ablations isolate the key ingredient. The first is the tree planner truncated to depth one. The second is an open-loop variant that scores test sequences by their expected value without conditioning later choices on earlier outcomes. Both share the planner's model, scenario set and compute.
All methods get the same training split, the same density model where applicable, and the same validation budget for tuning stop thresholds and cost weights. They are compared at matched per-patient wall-clock budgets.
Beating a baseline means a lower expected total cost at equal sensitivity for the high-risk class, with a bootstrap confidence interval that excludes zero. The area under the cost–accuracy curve is reported as the summary.
Why not just…?
Why not just use the silhouette model and threshold it? For the VAT target this may be nearly enough, given R² 0.885. We expect the planner's gain on that target to be small and confined to patients near the threshold, and we will report it that way. The VAT/ASAT ratio target, where silhouettes reach only about R² 0.5, is where a single cheap estimate leaves many patients genuinely uncertain. Comparing the two targets tests whether the method's value tracks cheap-test informativeness as predicted.
Why not just use ACO re-planned at each step? This is the real threat. Re-planning already makes ACO adaptive, and the paper shows non-parametric subset scoring beating learned sequential policies. Its weakness is specific. When scoring a subset like {silhouette, DXA}, it assumes both are bought. It therefore assigns the silhouette no value as a gate for DXA, and can skip straight to DXA or overbuy. The open-loop ablation uses this exact logic on our model, so the gap between it and the full tree measures option value directly. We will also vary the DXA-to-silhouette cost ratio. The prediction is that the gap widens as DXA becomes relatively more expensive and disappears when costs are similar. If it never appears, lookahead is not needed.
Why not just hand-tune a two-stage threshold rule? For a fixed three-test cascade, a tuned band rule may approximate the optimal policy. We include it for that reason. Where the rule should break is in adding a fourth test, imposing a radiation cap, or conditioning on sex. In those cases the number of thresholds to hand-tune grows and they interact. Whether the rule falls behind as tests are added is measured, not assumed.
Proposed Experiments
The primary testbed is the UK Biobank imaging cohort used by the silhouette study: 40,032 participants with whole-body MRI depot volumes and silhouettes derived from that MRI. Anthropometrics are added, and DXA is used for the subset that has it. The tests are evaluated by replay as described above. One caveat applies: the silhouettes come from the MRI itself, so we treat silhouette "cost" as that of a standalone optical or photographic capture. We flag this as an assumption that the Shape Up! replication addresses.
The secondary testbed is Shape Up! Adults from Tian et al.. There the cascade is anthropometrics, then a 3D optical scan with mesh-autoencoder plus Gaussian-process estimates, then DXA as truth. This tests transfer to a different cheap modality and gives a sex-stratified check where learned shape features are known to be weaker for women.
AbdCTBench offers only a surface-to-CT pair. We use it solely to stress-test the observation model on a single-site cohort, not as a cascade.
Decisive check 1. Run all policies across a sweep of cost ratios and radiation caps. Plot the cost–accuracy frontier for each. Test whether full-tree minus open-loop-ablation savings rise with the DXA/silhouette cost ratio and are largest on the ratio target. A flat, near-zero gap refutes the core claim.
Decisive check 2. Repeat with sex-blind and sex-conditioned observation models, and report per-sex calibration of predicted test outcomes and per-sex cost savings. The hypothesis is that a sex-blind planner over-trusts cheap scans for women and under-orders DXA, which shows up as lower sensitivity in females at equal cost.
Decisive check 3. Shrink the number of scenarios K and measure how close the sparse tree gets to a densely sampled reference planner. This shows how many sampled futures are needed for stable choices with continuous test outcomes.
Decisive check 4. Deliberately miscalibrate the density model and measure replayed performance. This quantifies how much model error costs a lookahead planner relative to the k-NN-based ACO, which makes weaker modelling assumptions.
References
- Cohort-Based Active Modality Acquisition
- Acquisition Conditioned Oracle for Nongreedy Active Feature Acquisition
- Silhouette images enable estimation of body fat distribution and associated cardiometabolic risk
- Association between cognitive impairment and comprehensive weight phenotypes among older adults in China: a nationwide survey
- 3D convolutional deep learning for nonlinear estimation of body composition from whole body morphology
- DESPOT: Online POMDP Planning with Regularization
- A Sparse Sampling Algorithm for Near-Optimal Planning in Large Markov Decision Processes
- AbdCTBench: Learning Clinical Biomarker Representations from Abdominal Surface Geometry