# Prediabetes Status Task Card Task id: `prediabetes_status` Evidence status: grounded Evidence reviewed: 2026-08-15 ## Target Participant-level binary classification of prediabetes status from the curated diabetes phenotype table. Implementation facts: - target dataset release: `anat_curated_phenotype` (the mixin default, `_curated_phenotype_status.py:62`) - target table: `anat.curated_phenotype.diabetes`, resolved from `anat_table = "diabetes"` - target column read: `curated_phenotype`. **This is hardcoded in the loader** (an internal review record), not taken from the Task - `target_field` on the class is `diabetes__curated_phenotype` and is **inert on this release** - it is read only by the alternate `frozen_v1_2` route (`_curated_phenotype_status.py:336`). Do not treat it as the operative field - label map: `{"non-diabetes": 0, "prediabetes": 1}` - excluded states: `diabetes`, `low confidence diabetes`, missing - default research stage: `00_00_visit` - default covariates: age, sex, BMI from baseline anthropometrics - primary metric: ROC AUC ## Why This Matters Prediabetes is a clinically important metabolic risk state and a distinct Yeela-selected DS-1584 condition. It is separate from `t2d_status`, which explicitly excludes prediabetes from its T2D-vs-normoglycemic definition. **What a row here does not support.** This is same-visit discrimination of a curated status label. It is not a diagnosis, not incident-prediabetes prediction, and not evidence that any feature set would change management. The label itself is a curation product, so a high AUC is a statement about recovering that curation, not about identifying a clinical state. ## Related HPP Measurements Demonstrated in HPP: - **Prediabetes is among the incident outcomes CGM range-occupancy predicts in this cohort.** Godneva et al. followed 6,550 HPP participants a mean of 2.6 ± 1.3 years; 444 reported a new metabolic condition (prediabetes, diabetes, MASLD or hyperlipidemia), and time above 140 mg/dL carried an age- and sex-adjusted HR of 1.34 (95% CI 1.26-1.42) (Godneva et al. 2026). That is *incident* risk on self-reported diagnoses, a different estimand from this Task's same-visit label. - **The fasting-glucose leg of any prediabetes definition is unstable in HPP.** Of 5,328 HPP participants normal on a first fasting-glucose reading, 40% fell into the prediabetes range on sequential measurement (Shilo et al. 2024). - CGMap gives the HPP reference distribution for CGM measures in non-diabetic individuals (Keshet et al. 2023). **The committed rows settle half of the Benchmark-Track hypothesis below, which an earlier version of this card said they did not.** On the logistic paired comparison against the demographic floor: V1 CGM **ΔAUROC = +0.049 (95% CI [+0.018, +0.081])** on n = 11,320 clears the floor; V1 DXA **+0.020 ([-0.004, +0.044])** on n = 8,868 does not. So CGM summaries discriminate this curated label beyond age/sex/BMI and DXA body composition is not demonstrated to. Proposed rather than shown: that either reflects prediabetes physiology rather than the curation's own inputs. ## Clinically Meaningful Variants The binary is already the categorical variant, so the live question is whether its boundary is the right one. **The ordinal label is the defensible extension; a threshold-derived variant is not.** ADA defines prediabetes by FPG 100-125 mg/dL, HbA1c 5.7-6.4%, or a 2-hour OGTT value (ADA 2026), and WHO's impaired-fasting-glucose band starts at 110 mg/dL instead - so "prediabetes" is not one measurement with one cutoff, and reconstructing it from HPP lab fields would mint a fourth definition rather than apply a validated one. The curated panel exists precisely to avoid that, and it combines self-report with glycemic support features rather than thresholding a single live lab. Two consequences worth stating plainly: - **The curated label is not interchangeable with an ADA or WHO classification**, and a row here must not be reported as ADA-defined prediabetes prevalence or detection. - **Excluded is not negative.** Curated `diabetes` and `low confidence diabetes` participants are dropped, not scored as controls, so the operating population is prediabetes-vs-non-diabetes and the positive rate is not a cohort prevalence. ## Benchmark-Track Evidence Benchmark-Track Evidence reviewed: 2026-07-29. literature review supports CGM summaries and DXA body composition as compact status-discrimination panels beyond age, sex, and BMI. CGM differences between prediabetes and normoglycaemia are cross-sectional correlates, not incident disease prediction. HbA1c, fasting glucose, diabetes-condition fields, and glucose-lowering medication must stay out of predictors because they support the curated label. ## Baselines And Ceilings The required floor is a no-modality-feature classifier baseline: `NoFeaturesPredictor` contributes no modality features, and `DemographicClassifierFloorStrategy` fits age, sex, and BMI. Useful future comparators include HbA1c, fasting glucose, CGM summaries, dietary features, gut microbiome species, and MMFM participant embeddings. ## Caveats - Labels come from the curated diabetes panel, which combines self-report and glycemic support features rather than a single live lab threshold. - Participants with curated diabetes or low-confidence diabetes are excluded, not treated as controls. - The diabetes panel is forward-propagated by the upstream curation. - HbA1c and fasting glucose are close to the diagnostic boundary and should be interpreted as strong clinical comparators, not neutral features. ## Metrics Primary: ROC AUC. Secondary: average precision, accuracy, balanced accuracy, Brier score, positive rate, split sizes. ## References - DS-1584 Yeela-selected condition curation: `pha_condition_projection/DS-1584/condition_candidates_detailed.csv`. **Does not resolve** in this repository or in HPP documentation / HPP documentation / HPP documentation; kept because it is the only pointer to how this condition was selected, and marked so it is not read as a repo path. - Existing `t2d_status` Task Card for the T2D-vs-normoglycemic boundary. - HPP `curated_phenotypes` dataset docs for the diabetes curated phenotype. - Rákóczi et al. (2026), CGM classification versus HbA1c and fasting glucose, . - Godneva et al. (2026), the spectrum of normoglycemia and time in glycemic ranges, *Diabetes Care*, . - Shilo et al. (2024), continuous glucose monitoring and intrapersonal variability in fasting glucose, *Nature Medicine*, . - Keshet et al. (2023), CGMap, *Cell Metabolism*, . - ADA Standards of Care in Diabetes 2026, diagnosis and classification, . ## Open Questions - Should a future Task expose the full non-diabetes / prediabetes / diabetes ordinal label instead of this binary prediabetes-vs-non-diabetes slice? - What exactly does the curated panel use to assign `prediabetes`, and how much of it is self-report? The card can say the panel combines self-report with glycemic support features, but no committed artifact in this repo records the rule, so the label's relationship to ADA or WHO criteria is undocumented on our side. Until that is written down, the non-claim above is doing load-bearing work.