# HbA1c Task Card Task id: `hba1c` Evidence status: grounded Evidence reviewed: 2026-08-15 ## Target Cross-sectional HbA1c regression on HPP V1. Implementation facts: - target field: `bt__hba1c` - target unit: percent (NGSP/DCCT units, not IFCC mmol/mol) - target dataset: `blood_tests` - target source: `weizmann` - default research stage: `00_00_visit` - valid range: 3.0-15.0 percent; a PhenoBench plausibility filter, not a diagnostic interval - primary metric: R2 ## HPP data source Target loads from the pinned Weizmann `blood_tests` source. The live operational source HMO table was audited as a candidate replacement. **Audit decision: keep Weizmann.** Paired agreement is strong (94.9% exact), but operational source + `events` yields net −3.5% baseline coverage with ~23% cohort turnover and no offsetting gain, and 998 operational source participants carry hba1c with no recorded unit (`%` vs `mmol/mol`). Rationale and reconsideration triggers: blood-test source audit §9. ## Why This Matters HbA1c is the standard long-horizon glycemic marker. It captures average glucose exposure over weeks to months and is clinically tied to diabetes diagnosis, diabetes progression, microvascular complications, and cardiovascular risk. For multimodal foundation-model evaluation, HbA1c is a direct metabolic target: CGM-derived signals should have a plausible route to predicting it. That makes it useful as an early sanity check for whether an embedding contains metabolic information beyond demographics. **What a row here does not support.** A predicted HbA1c is a same-visit association with one laboratory value. It is not a diagnosis, does not establish glycemic control over any future period, and says nothing about microvascular or cardiovascular outcomes - those are the endpoints HbA1c is *validated against* in the clinical literature, not endpoints this Task measures. ## Related HPP Measurements Demonstrated in HPP, and one of these should temper the expectation above: - **In this cohort's non-diabetic range, CGM time-in-range and HbA1c are almost unrelated.** Godneva et al., 8,687 HPP adults, report TIR correlating with HbA1c at r = 0.04, with time in tight range showing no significant association at all - while the same TIR correlates with mean glucose at r = 0.62 and with the glucose management indicator at r = 0.63 (Godneva et al. 2026). So "CGM should predict HbA1c" holds for *mean-glucose* summaries far better than for range-occupancy summaries, and the restricted non-diabetic range is doing much of the work. - Time above 140 mg/dL carries prospective signal for incident metabolic conditions (age- and sex-adjusted HR 1.34, 95% CI 1.26-1.42), which HbA1c itself is not shown to capture here (Godneva et al. 2026). - CGMap gives the HPP reference distribution of CGM-derived measures in non-diabetic individuals (Keshet et al. 2023). **A matched-cohort comparison is already committed, and an earlier version of this card said none was.** On the snapshot's `Matched-cohort comparison` track, n = 1,103, the full IGLU panel reaches **ΔR2 = +0.105 (95% CI [+0.010, +0.202])** over the demographic floor, while Nightingale metabolites (+0.005 [-0.030, +0.035]), DXA regional lean mass and grip strength all span zero. So CGM summaries clear the floor on matched support and the other modalities do not - which is the comparison this card asked for. Read that against the Godneva reading above rather than instead of it: the IGLU panel is mean-glucose-like, and it is range-occupancy that fails to track HbA1c. Proposed rather than shown: that a *learned* CGM representation beats the engineered panel. The physiological-time-series ECG route below is the one place a clearly negative answer is already recorded. ## Clinically Meaningful Variants **A categorical HbA1c Task is more defensible than a categorical FPG one, and still not defensible on this measurement as configured.** ADA classifies HbA1c ≥6.5% as diabetes and 5.7-6.4% as prediabetes (ADA 2026), and unlike a fasting draw HbA1c integrates weeks of exposure, so it is far less sensitive to day-of-visit state. Three things block a threshold Task here anyway: - **Units and assay alignment are not incidental.** The target is in NGSP/DCCT percent; IFCC mmol/mol is the other reporting scale, and a threshold applied to the wrong scale is silently wrong rather than loudly wrong. The blood-test source audit already found 998 operational source participants carrying an hba1c value with **no recorded unit**, which is exactly this failure waiting to happen on a source swap. - **HbA1c is not a pure glucose readout.** Anything shortening red-cell lifespan biases it downward - modelled and verified for T2DM patients by Zhang et al. (2025) - and haemoglobin variants shift it independently of glycemia, with Kerdsinchai et al. (2025) reporting significantly lower HbA1c in non-diabetic haemoglobin H disease and calling for disease-specific reference ranges. A binary Task inherits that bias as label error concentrated in a subpopulation, which a continuous target merely blurs. - The cohort is largely non-diabetic by construction, so a ≥6.5% positive class would be small and the boundary class (5.7-6.4%) would dominate - the same class-degeneracy hazard `t2d_status` documents for hash splits. The continuous target stays primary. `prediabetes_status` and `t2d_status` are the supported categorical glycemic routes, and both take labels from the curated diabetes phenotype rather than from a threshold on this field. ## Baselines And Ceilings The required floor is a no-modality-feature baseline: `NoFeaturesPredictor` contributes no modality features, and `DemographicFloorStrategy` fits age, sex, and BMI. Meaningful model claims should clear that floor on the same cohort and split. For CGM-adjacent models, simple glucose-summary or IGLU-feature baselines are strong comparators, because HbA1c is closely tied to mean glucose. Expected demographic floor from the source material is roughly R2 0.10-0.18, mostly age-driven. Empirical HPP floors should be treated as the authority once available. ## Physiological Time-Series V1 the internal evaluation configuration evaluates a residual 1-D CNN trained end-to-end on target-blind samples from the baseline 10-second, 12-lead ECG. It is a same-visit representation probe inspired by published raw-ECG HbA1c prediction, not forecasting and not a substitute for laboratory HbA1c. The primary result is held-out delta-R2 versus the age/sex/BMI floor on identical support. Held-out exploratory result (test n=343): temporal R² 0.0127 versus tuned age/sex/BMI R² 0.1143, delta-R² -0.1016 (95% CI [-0.1730, -0.0390]); temporal MAE 0.3237% versus 0.2983%, delta-MAE +0.0254% (95% CI [+0.0128, +0.0385]). This route did not beat demographics. ## Caveats - HbA1c is not pure glucose exposure. RBC lifespan, hemoglobinopathies, anemia, and other non-glycemic biology can shift it. - Medication changes can dominate longitudinal HbA1c deltas. - A model that mainly learns age, sex, or BMI is not clinically interesting for this task, even if raw R2 looks acceptable. - CGM-derived features have a strong path to HbA1c, so failure here is a useful negative signal for metabolic encoders. - Conversely, success here alone is not enough to claim broad metabolic understanding. ## Metrics - Primary: R2. - Secondary: MAE. - Useful audit views: residualized performance after demographics, split Ns, cohort ID, and comparison to CGM-feature baselines. ## References - Source material: `docs/task_cards/source_material/foundation-model-surrogate-targets-v1.md` - UK Biobank HbA1c field 30750: - HbA1c and cardiovascular disease meta-analysis: - Godneva et al. (2026), the spectrum of normoglycemia and time in glycemic ranges, *Diabetes Care*, . Source of the TIR-HbA1c r = 0.04 reading above. - Keshet et al. (2023), CGMap, *Cell Metabolism*, . - ADA Standards of Care in Diabetes 2026, diagnosis and classification, . - Zhang et al. (2025), rectifying the impact of shorter red blood cell lifespan on HbA1c, *Frontiers in Endocrinology*, . - Kerdsinchai et al. (2025), HbA1c levels in haemoglobin H disease, *Biochemistry and Biophysics Reports*, . ## Open Questions - What HPP empirical floor should become the canonical threshold for "clears demographics"? - Which CGM-feature baseline is the right near-term ceiling: IGLU, mean glucose, or a broader engineered-feature model? The Godneva reading argues for a mean-glucose summary over a range-occupancy one, but that is an inference from a correlation table, not a measured PhenoBench comparison. - Should longitudinal HbA1c delta become a separate task card rather than a variant of this one? - Does HPP record haemoglobinopathy or anaemia status well enough to flag the participants whose HbA1c is biased for non-glycemic reasons? Unresolved, and it bounds how far a residual analysis can go.