# eGFR Task Card Task id: `egfr` Evidence status: grounded Evidence reviewed: 2026-07-29 ## Target Cross-sectional eGFR regression on HPP V1 using the CKD-EPI 2021 race-free formula. Current PhenoBench implementation computes eGFR from: - serum creatinine: `bt__creatinine` - age - sex Implementation facts: - target name: `egfr` - target field: `egfr` - target dataset: `derived_renal_indices` - target source: **`derived`**, not `weizmann`. This is a computed target, and `target_source` is what the Benchmark Track sees. The Weizmann `blood_tests` pin describes where the **creatinine input** comes from - see HPP data source below - and naming it here misdescribed the declared source - default research stage: `00_00_visit` - valid range: 10.0-150.0 mL/min/1.73m2 - primary metric: R2 ## HPP data source The creatinine input loads from the pinned Weizmann `blood_tests` source. The live operational source HMO table was audited as a candidate replacement. **Audit decision: keep Weizmann.** Creatinine agreement is tight, but the CKD-EPI formula amplifies per-participant differences (eGFR p99 |Δ| ≈ 10.6), and operational source + `events` yields net −18.1% baseline coverage with ~34% cohort turnover. Rationale and reconsideration triggers: blood-test source audit §9. ## Why This Matters eGFR is the standard clinical estimate of kidney filtration. It is used for CKD staging, medication dosing, kidney-disease monitoring, and risk stratification. It is a clinically recognizable target and connects to long-term kidney and cardiovascular outcomes. For PhenoBench, eGFR is also a useful stress test for benchmark interpretation: the target is formula-derived, and the formula already includes age and sex. ## Baselines And Ceilings The required floor is a no-modality-feature baseline: `NoFeaturesPredictor` contributes no modality features, and `DemographicFloorStrategy` fits age, sex, and BMI. This floor is structurally inflated because age and sex are inputs to CKD-EPI. Current code documents Surrogate's eGFR demographic floor around R2 0.22. The source material estimates an age+sex floor around R2 0.15-0.25. The remaining variance is largely creatinine-driven. Creatinine reflects kidney function, muscle mass, diet, and assay effects. A model with direct access to Nightingale creatinine or closely related chemistry can trivially look strong. ## Caveats - Age and sex are directly embedded in the target formula. - High raw R2 can mean the model recovered formula inputs, not kidney-specific physiology. - Creatinine is not a pure kidney-function marker; it also reflects muscle mass and diet. - Nightingale creatinine and clinical serum creatinine may differ by aliquot and platform, creating systematic offsets. - Single-timepoint eGFR does not fully establish chronic kidney disease, which clinically requires chronicity. ## Metrics - Primary: R2 for continuous eGFR regression. - Secondary: MAE. - Possible extension: AUROC or ordinal metrics for CKD categories, with clear chronicity caveats. - Essential audit: residualized analysis that separates formula-input recovery from kidney-specific signal. ## Benchmark-Track Evidence literature review and broader literature search support V1 Nightingale as the minimal cross-modal track. Large population NMR studies associate lipoprotein, fatty-acid, and GlycA traits with creatinine-based eGFR after demographic adjustment, but the relationship is cross-sectional and may reflect reduced clearance or altered metabolism. The benchmark must exclude creatinine and all eGFR aliases. DXA was not selected: lean mass changes creatinine generation and would make this formula-derived target especially difficult to interpret. ## References - Source material: `docs/task_cards/source_material/foundation-model-surrogate-targets-v1.md` - UK Biobank creatinine field: - UK Biobank cystatin C field: - Kidney function trajectories from UK Biobank EMR: - Barrios et al. (2018), NMR biomarkers of renal function: . - Nankivell et al. (2020), muscle mass and creatinine eGFR: . ## Open Questions - Should a creatinine-residualized eGFR variant be a separate task? - Should CKD staging become a curated phenotype card rather than an eGFR regression variant? - **Resolved (#6, 2026-08-20): it does not warn, it refuses or corrects.** Where dropping the overlapping covariate leaves a demographic floor, the config drops it (FLI loses BMI, FIB-4 loses age) and the row keeps competing. eGFR is the case where it does not: CKD-EPI takes creatinine, age AND sex, so removing the overlap would leave a BMI-only floor and model arms stripped of age and sex. So eGFR's rows keep their covariates, declare `covariate_in_formula: [age, sex]`, and carry `benchmark_role: reference` instead of ranking. Model-accessible biomarkers are a separate question and this does not answer it: creatinine access is excluded by the panel rather than by this mechanism.