# FIB-4 / FIB-3 Task Card Task ids: `fib4`, `fib3` Evidence status: grounded Evidence reviewed: 2026-07-28 Benchmark status: complete ## Target Cross-sectional FIB liver-score regression on HPP V1. The shared card covers two related derived targets: - `fib4`: the standard age-containing FIB-4 score. - `fib3`: the age-residualized sibling without the explicit age term. Current PhenoBench implementation computes: ```text FIB-4 = age * AST / (platelets * sqrt(ALT)) FIB-3 = AST / (platelets * sqrt(ALT)) ``` Implementation facts: Shared by both Task ids: - source components: age, AST, ALT, platelets - AST field: `bt__ast_got` - ALT field: `bt__alt_gpt` - platelet field: `bt__platelets` - target source: `weizmann` - default research stage: `00_00_visit` - primary metric: R2 For `fib4`: - valid range: 0.1-10.0 For `fib3`: - valid range: 0.001-1.0 ## HPP data source The AST / ALT / platelet inputs load from the pinned Weizmann `blood_tests` source. The live operational source HMO table was audited as a candidate replacement. **Audit decision: keep Weizmann.** operational source + `events` yields net −23.9% baseline coverage with ~40% cohort turnover, low exact agreement (58.1%), and a platelet input that shows an html_parser–associated discrepancy. Rationale and reconsideration triggers: blood-test source audit §9. ## Why This Matters FIB-4 is a simple clinical liver-fibrosis score. It is widely used because it requires only age and routine blood tests. It has reported associations with cardiovascular events, cardiovascular mortality, and all-cause mortality. For PhenoBench, FIB-4 is useful partly because it is clinically familiar and partly because it exposes a common benchmark trap: formula-derived targets can look predictable for reasons that are not biologically interesting. ## Baselines And Ceilings The required floor is a no-modality-feature baseline: `NoFeaturesPredictor` contributes no modality features. **This card covers two Task ids and they now differ:** FIB-4's `DemographicFloorStrategy` fits **sex and BMI**, and FIB-3's still fits age, sex and BMI. Age was dropped from FIB-4 in #6 (2026-08-20) because it appears directly in the target formula's numerator, which made the floor structurally inflated: current code documents Surrogate's FIB-4 floor at about R2 0.31, while the source material estimates about R2 0.40-0.60 because age dominates the score. FIB-3 kept all three, because its formula takes AST, ALT and platelets and none of those is a floor covariate. **The FIB-4 scores below predate that change**, measured against the age/sex/BMI floor, so read them as historical rather than as the current matrix - figure 1 marks those contracts `historical` and their replacements `configured` with no verified route yet. **The FIB-3 history is unaffected.** The more meaningful liver-signal question is whether a model captures AST, ALT, and platelet-related information beyond age. `Fib3Task` is the cleaner paired task because it removes the explicit age term from the formula and has a much lower demographic floor around R2 0.07 in the task code. ## Benchmark-Track Evidence literature review supports Nightingale metabolomics as the primary systemic fibrosis modality and DXA body composition as secondary. The proposed fixed 27-field Nightingale panel contains no AST, ALT, platelets, or albumin, so it does not directly expose a FIB component. An age/sex/BMI floor is formula-circular for FIB-4 because age is a direct numerator term, which is why the v0 FIB-4 floor no longer holds age. FIB-3 should still accompany FIB-4 as the cleaner liver-signal headline; a strong FIB-4 score alone is not evidence of fibrosis biology. ## V0 Benchmark Execution The frozen same-visit V1→V1 matrices completed on 2026-07-29. All twelve rows across FIB-3 and FIB-4 are locally successful and remotely `FINISHED` in operational source. - FIB-3 Nightingale: n=4,371; test=656; best R2=0.123 (CPU TabSwift). DXA: n=2,908; test=443; best R2=0.105 (CPU TabSwift). - FIB-4 Nightingale: n=4,371; test=656; best R2=0.345 (Ridge). DXA: n=2,907; test=443; best R2=0.357 (CPU TabSwift). - FIB-4 demographic floors were R2 0.315-0.331, confirming the predeclared age-formula circularity; FIB-3 remains the cleaner sensitivity analysis. - Full rows and provenance: FIB-3 deep dive, FIB-4 deep dive, and aggregate review. ## Caveats - Age is literally in the FIB-4 formula, so demographic performance can be misleading. - ALT and platelets also vary with age, compounding the age signal. - FIB-4 was developed in diseased cohorts; interpretation in a healthy population is less straightforward. - Standard clinical cutoffs can overcall fibrosis in older healthy adults. - Strong raw R2 may reflect formula reconstruction, not liver-specific biology. ## Metrics - Primary: R2 for continuous score regression. - Useful companion: AUROC for clinically relevant FIB-4 thresholds, if the task is extended to classification. - Essential audit: report `Fib3Task` or age-residualized comparison whenever making a liver-specific claim. ## References - Source material: `docs/task_cards/source_material/foundation-model-surrogate-targets-v1.md` - Liver fibrosis scores prognosis meta-analysis: - Caussy et al. (2019), serum metabolites and advanced fibrosis: - Ciardullo et al. (2022), body-fat distribution, steatosis, and fibrosis: - Age and sex reference values for FIB-4 in healthy adults: - FIB-3 without age factor: - Demographic floor estimate source: ## Open Questions - Should FIB-4 remain a primary task, or should FIB-3 become the preferred liver signal task? - Which FIB-4 classification threshold is appropriate for a healthy HPP-like cohort? - **Resolved (#6, 2026-08-20): no warning.** The overlap is removed where the remainder is still a demographic floor, and the row stops competing where it is not. FIB-4 declares `demographic_features: [sex, bmi]`, dropping the age term the formula takes. A row that still supplies its target's ingredients carries `benchmark_role: reference` and `test_a_publishing_row_that_declares_a_formula_overlap_does_not_rank` refuses the ranked combination, so the guarantee is a refusal rather than a label. **FIB-3 is unchanged and keeps age, sex and BMI**, because its formula takes AST, ALT and platelets and none of those is a floor covariate, so it had no overlap to remove. This card is shared between the two Task ids, so the distinction matters: the floor comparison between them is no longer like-for-like on covariates, which is the point rather than a side effect.