# CIMT Task Card Task id: `cimt` Evidence status: grounded Evidence reviewed: 2026-08-15 ## Target Mean carotid intima-media thickness from HPP carotid ultrasound. Continuous-scalar regression. The default research stage is `00_00_visit`, but the Task can be configured for another available stage. Implementation facts: - target field: `imt_mean` - target dataset: `carotid_ultrasound` - target source: None - target unit: mm - default research stage: `00_00_visit` - valid range: 0.3-3.0 mm; a PhenoBench plausibility filter, not a clinical interval - primary metric: R2 ## Why This Matters CIMT is a subclinical vascular-structure marker associated with atherosclerosis and cardiovascular risk. It is clinically interpretable but not a current headline screening target, so PhenoBench should treat it as a secondary vascular-aging target. **What a row here does not support.** A predicted CIMT is a same-visit association with an operator-acquired ultrasound measurement. It is not a diagnosis of atherosclerosis, not a plaque assessment, and not a cardiovascular risk estimate. Guidelines have moved away from CIMT as a risk-stratification tool precisely because adding it to established risk scores did not improve prediction usefully, so a strong row here is a statement about the measurement, not about risk. ## Related HPP Measurements Demonstrated in HPP and persisted in this repository, from an internal review record, test split: - V1 fundus + demographics reaches test R2 = 0.299 (95% CI [0.256, 0.342]) on a cohort of 7,403. - V1 Nightingale + demographics reaches R2 = 0.270 (95% CI [0.224, 0.310]) on 6,983. **The paired comparison is committed and it confirms this directly**: fundus reaches **ΔR2 = +0.015 (95% CI [-0.002, +0.030])** and Nightingale **+0.006 ([-0.012, +0.023])** - both span zero. So neither track demonstrates vascular-morphology signal beyond age, sex and BMI on this cohort. **That is the finding to carry**: CIMT is strongly age-dependent, a raw R2 near 0.3 is mostly age being recovered, and the measured margins are indistinguishable from zero. Retinal microvasculature leading again matches the ordering on `sbp` and `hypertension_status`, where fundus also leads - three targets on one axis with the same modality ranking. **These are `reviewable` rows, not `final` ones.** Measured across the committed leaderboards: all 787 rows carry `curation_state: reviewable` and none is `final`. A complete fresh v0 re-run is committed to, so every figure in this section is pre-re-run evidence that the re-run supersedes - persisted and citable today, not a frozen result. **Which curation states the Figure 1 export accepts is the export track's live question, not this card's**; read the two counts above, not a gate. ## Clinically Meaningful Variants **No categorical CIMT variant is defensible, and unlike the lipid targets the problem is that the measurement itself is not standardised across devices.** CIMT values depend on the scanner, the insonation angle, which carotid segments are measured, whether plaque is included, and the reading convention (mean of means, maximum, near or far wall). Published "high CIMT" cutoffs and percentile-based abnormality rules are therefore protocol-specific, and the common clinical use is a **percentile for age and sex within the measuring laboratory's own reference**, not a fixed millimetre threshold. Two consequences: - **A fixed millimetre bin would encode this laboratory's protocol as though it were a clinical standard.** The card's `imt_mean` is one convention among several, and nothing here establishes agreement with the references those cutoffs come from. - **An age-sex percentile variant would be circular on this Task.** Since most of the predictable variance is age, a percentile-for-age label removes the signal the demographic floor already captures and leaves a much harder, smaller-variance target - which may be the more interesting Task, but it is a different one with its own floor. The continuous target stays primary. Carotid plaque presence, which the card lists as a comparator, is the categorical vascular variant with a defensible definition, and it is a separate measurement rather than a bin on this one. ## Baselines And Ceilings The required floor is a no-modality-feature baseline: `NoFeaturesPredictor` contributes no modality features, and `DemographicFloorStrategy` fits age, sex, and BMI. The task code records a moderate demographic floor around R2 0.28 because CIMT is strongly age-dependent. Natural comparators include carotid plaque, SBP, lipid markers, retinal vessel features, ECG/vascular embeddings, and clinical risk-factor models. Claims should focus on vascular morphology signal beyond age. ## Caveats - Age can dominate raw performance. - CIMT adds modest incremental risk information beyond standard clinical risk factors in some settings. - Ultrasound measurement protocol and segmentation quality can add noise. - Strong CIMT prediction should not be claimed as incident CVD prediction unless evaluated against event outcomes. ## Metrics Primary metric is R2. Secondary MAE in mm and Spearman are useful. Audit views should include demographic-floor delta R2 and age-stratified residuals. ## References - Source material: `docs/task_cards/source_material/foundation-model-surrogate-targets-v1.md` - Tschiderer et al., 2020, CIMT and incident carotid plaque meta-analysis: - Mitra et al., 2025, CIMT and CVD risk factors in UK Biobank: - Talmor-Barkan et al. (2026), RetiMap automated retinal vascular measurement. **Manuscript in preparation, no DOI** - treat as HPP context rather than a citable source. Local copy in **HPP documentation**: `assets/papers/hpp/talmor_barkan_2026_retimap_automated_retinal_vascular_measu/`. ## Benchmark-Track Evidence Benchmark-Track Evidence reviewed: 2026-07-29. NMR metabolomics and retinal morphology are the proposed V1 systemic tracks; same-visit carotid, BP, and vascular features are excluded. CIMT remains V1-only until a like-for-like V2 target is audited. See `04-cardiovascular-lipid.md`. ## v0 Final-Test Status The classical release is complete: Ridge/logistic and GBDT are remotely verified on every approved exact cohort and rendered in the Figure 1 task deep dives. ## Open Questions - Should carotid plaque or vascular age become a stronger vascular endpoint than CIMT? - What ultrasound QC fields should be surfaced in run review packets? - Should CIMT rows be excluded from headline demos unless age-residualized performance is reported?