# CGM Coefficient of Variation Task Card Task id: `cgm_coefficient_of_variation` Evidence status: grounded Evidence reviewed: 2026-07-29 Benchmark status: complete (V1 demographic floor; cross-modal value-add not run — #28) ## Target CGM Coefficient of Variation as a continuous HPP scalar. Implementation facts: - bodily system: metabolic/metabolomics - target route: `cgm.iglu_cv` - unit: % - acquisition: Abbott Libre Pro/Pro IQ CGM, nominally 14 days - construction/QC: upstream 100 × SD / mean over QC-passing connection, then participant-stage mean - valid range: 0-100% - paper fidelity: paper-inspired scalarization of HealthFormer's raw CGM modality ## Why This Matters CGM CV is a scale-normalized glycaemic-variability measure. This is a V1 cross-sectional demographic-floor benchmark (cross-modal value-add tracked in #28), not forecasting. Prediction does not establish instability-related clinical risk or diabetes status. ## Benchmark-Track Evidence literature review, CGMap, and HPP insulin-resistance work support DXA body composition and a non-glycaemic Nightingale panel as compact independent cross-modal tracks. Evidence is mainly cross-sectional association; the benchmark tests held-out prediction. Every same-visit CGM field remains masked. ## Related HPP Measurements CGM mean/SD/MAGE/CONGA, TIR, GMI, HbA1c, meals, activity, and medication. ## Clinically Meaningful Variants The continuous measurement is primary. The international consensus target of CV <36% distinguishes stable from unstable glycemia and was set from hypoglycemia risk in diabetes management (Danne et al. 2017; Battelino et al. 2019); healthy non-diabetic adults sit well below it (within-person CV about 17%). That threshold is a glycemic-management target, not a diagnosis, and it is derived largely in diabetes populations, so it does not define a label in this predominantly non-diabetic HPP Task. ## Baselines And Ceilings V1-only evaluation compares age/sex/BMI linear and tree floors on one cohort/split. LOCF and prior-target baselines are not run for this wave because paired V1/V2 coverage is inadequate. ## Caveats - The planned cross-modal shape (#28) must mask the full same-visit CGM family and raw stream, explicitly including mean and SD components that reconstruct CV; this wave runs only the demographic floor. - Only 222 paired V1/V2 summaries exist; this Task is V1 same-visit only. CV is unstable when mean or usable wear duration is poor, so retain upstream QC; note CV can also rise even as variability improves if mean glucose falls. - A good score supports prediction of this measurement only; it does not establish diagnosis, causality, or intervention response. ## Metrics Primary R2. Also report RMSE, MAE, Pearson r, Spearman rho, and split Ns. Target-distribution (quantiles), constant/median-baseline, and ceiling diagnostics are not yet computed for this wave; they are tracked in #35. ## References - Lutsker et al. (2026), HealthFormer, . - Keshet et al. (2023), CGMap HPP CGM atlas, . - Battelino et al. (2019), CGM time-in-range consensus, . - Broll et al. (2021), iglu R package for CGM metrics, . - Danne et al. (2017), international consensus on CGM (CV <36% target), . - Shah et al. (2019), CGM profiles in healthy people without diabetes (within-person CV ~17%), . - Selvin et al. (2023), within-person and between-sensor CGM metric variability, . - HPP documentation `cgm` dataset documentation and QC inventory. ## Execution Evidence - protocol: HPP V1 cross-sectional; V2 repeat support was insufficient - common cohort: n=11483; train=7906, validation=1832, test=1745 | Baseline | Strategy | Test R2 | RMSE | Pearson r | | --- | --- | ---: | ---: | ---: | | Demographic | `demographic_floor` | 0.020 | 4.023 | 0.142 | | Trees | `demographic_tree` | 0.021 | 4.019 | 0.151 | ## Open Questions - Domain-expert review remains outstanding; evidence grounding and real-run gates are complete.