# VAT Task Card Task id: `vat` Evidence status: grounded Evidence reviewed: 2026-07-29 ## Target DXA-derived visceral adipose tissue mass. Continuous-scalar regression. The default research stage is `00_00_visit`, but the Task can be configured for another available stage. Current PhenoBench implementation: - target field: `total_scan_vat_mass` - target dataset: `dxa` - default research stage: `00_00_visit` - valid range: 1.0-6000.0 g - primary metric: R2 ## Why This Matters VAT is a central adiposity readout for metabolic risk. Unlike BMI, it measures visceral fat directly and captures a body-composition axis that routine screening can miss. The Shilo 2026 insulin-resistance surrogate paper makes VAT a useful paper-to-PhenoBench pilot because its headline result maps to the existing `vat` Task rather than a new target. The paper predicts DXA VAT from routine clinical measurements, so the PB asset mapping is: - Task: existing `vat` - Predictor: `hpp_participant_features` - Strategy: `ridge` - Configs: `vat_anthropometric_routine.*` and `vat_circumference_routine.*` ## Baselines And Ceilings - **Floor**: no-modality-feature baseline. `NoFeaturesPredictor` contributes no modality features, and `DemographicFloorStrategy` fits age, sex, and BMI. - **Anthropometric routine track**: `HppParticipantFeaturePredictor x RidgeStrategy`, adding weight, height, sitting systolic BP, and sitting diastolic BP to the task covariates. This represents Shilo's paper feature set after keeping age, sex, and BMI in the task covariate channel. - **Circumference routine track**: same as the anthropometric routine track, plus waist circumference, hip circumference, and waist-to-hip ratio. - **Engineered gait track**: `HppModalityFeaturePredictor x RidgeCVStrategy` over `gait_movements` engineered features. This is a Gabet-2026-inspired analogue against the existing VAT target, not the paper's Gait Fusion embedding result. The `vat_gait_engineered_visit2.*` row pins the target and predictor to `02_00_visit` because the baseline visit has weak gait/DXA overlap for this paper-shaped row. - **Fundus AutoMorph track**: `FundusAutoMorphFeaturePredictor x RidgeCVStrategy` over the 18 loader-native AutoMorph vessel features. This is a RetiMap-inspired cross-sectional analogue against VAT, not the paper's incident-CVD, survival, or paper-specific feature-cluster result. - **Sleep summary track**: `HppParticipantFeaturePredictor x RidgeCVStrategy` over loader-native WatchPAT respiratory, architecture, heart-rate, and SpO2 summaries averaged across baseline nights. This is a Kohn-sleep-inspired analogue, not the paper's full PRV/HRV or sex-stratified phenome-wide model. - **Gut microbiome species track**: `GutMicrobiomeAbundancePredictor x RidgeCVStrategy` over HPP `gut_microbiome` MetaPhlAn4 species features. This is a Rothschild/Keshet/Zahavi-shaped raw-abundance analogue, not gut-microbiome embedding parity or a disease-atlas association scan. - **CGM IGLU track**: `CgmIgluPredictor x RidgeCVStrategy` over stable-loader CGM summary features. This is a Shilo/GluFormer-shaped analogue against VAT, not GluFormer embedding parity or a paper-faithful insulin-resistance model. ## Benchmark-Track Evidence literature review and literature search identified a replicated, BMI-independent plasma NMR signature of VAT. A separate external prediction study found little added Nightingale value after waist was supplied; the v0 track therefore excludes waist and every DXA field and asks for lift over only age, sex, and BMI. Existing gait, CGM, sleep, microbiome, and routine-feature examples remain exploratory rather than additional frozen tracks. ## Caveats - DXA failed-scan zeros must not become biological low-VAT labels. The task's valid range and loader path should keep filtering obvious invalid values. - BMI is already a task covariate. Do not duplicate BMI in the predictor output for this Shilo mapping unless the task covariate contract changes. - Paper R2 values are not strict parity targets because PhenoBench uses canonical participant splits. - The Shilo paper reports R2 and Spearman rho. Scalar tasks now log Spearman as a secondary metric, but R2 remains the primary leaderboard metric. - Gabet's paper-faithful Gait Fusion embeddings/raw skeleton route is not exposed through the HPP loader. `vat_gait_engineered.*` rows must be described as engineered-gait analogue rows. - Gabet stage-matched rows must report their visit explicitly. The `04_00_visit` route was tested and should not be promoted from the initial evidence because the held-out R2 was unstable despite a positive rank correlation. - `vat_fundus_automorph.*` rows must keep only good-quality central-position fundus captures and participant-mean AutoMorph vessel features. Disc/cup features, image paths, QC scores, side/position metadata, and incident-CVD outcomes are out of scope. - `vat_sleep_summary.*` rows may include AHI/RDI/ODI as sleep predictors for VAT, but those fields would be direct leakage for an AHI target row. Do not reuse this feature list for `ahi` without an explicit leakage exclusion. - `vat_gut_microbiome_species.*` rows use the HPP loader's MetaPhlAn4 species frame, prevalence filtering, and canonical PB splits. They should not be compared as exact parity against papers using URS species, strain-level, pangenome, HUMAnN3 pathway, or learned embedding surfaces. - `vat_cgm_iglu.*` rows are same-visit CGM-summary analogues. They are useful for a CGM-metabolic axis benchmark, but they should not be read as causal, incident, GluFormer, or exact Shilo insulin-resistance surrogate parity. ## Metrics Primary metric is R2. Secondary metrics should include MAE in grams and Spearman rank correlation when available. For paper-inspired rows, report demographic-floor delta R2 and support-matched floor rows before making modality-value claims. ## References - Shilo et al. 2026, "Heterogeneity of Insulin Resistance Surrogates in Non-Diabetic Adults". - Shilo et al. (2026), heterogeneity of insulin-resistance surrogates, *medRxiv*, . Local copy in **HPP documentation**: `assets/papers/hpp/shilo_2026_heterogeneity_of_insulin_resistance_surr/`. - Neeland et al. (2019), replicated NMR signature of VAT, . - Boone et al. (2022), external VAT prediction comparison, . - Gabet et al. (2026), a gait foundation model predicting multi-system traits, *arXiv*, . Local copy in **HPP documentation**: `assets/papers/hpp/gabet_2026_a_gait_foundation_model_predicts_multi_s/`. - RetiMap mapping evidence in HPP documentation: `research_runs/phenobench-paper-task/next_batch_2026_06_22/`. - Kohn sleep mapping evidence in HPP documentation: `research_runs/phenobench-paper-task/next_batch_2026_06_22/`. - Gut microbiome mapping evidence in HPP documentation: `research_runs/phenobench-paper-task/hpp_full_corpus_asset_map_2026_06_22/` and prior `paper-replicate` runs for Rothschild, Keshet, Weissglas-Volkov, Zahavi, and Segev. ## Open Questions - Should VAT rows require support-matched demographic floors by default? - Which DXA QC fields should be exposed in review packets? - Should stage-matched gait/VAT rows remain paper-specific examples or become normal starter configs?