# Chronological Age Task Card Task id: `chronological_age` Evidence status: grounded Evidence reviewed: 2026-08-15 ## Target Participant chronological age. Continuous-scalar regression. The default research stage is `00_00_visit`, but the Task can be configured for another available stage. Implementation facts: - target field: `age` - target dataset: `anthropometrics` - target source: None - target unit: years - default research stage: `00_00_visit` - valid range: 18.0-120.0 years - primary metric: R2 - default covariates: sex, BMI - derived outputs: `biological_age_residual_years` and `biological_age_centered_residual_years` in `predictions.parquet` when predictions are persisted ## Why This Matters Chronological-age prediction is the safe PhenoBench surface for aging-clock papers. In Reicher/Bar 2024, biological-age scores are derived from held-out age predictions centered within age bins, so they should be logged as derived strategy outputs or artifact metadata unless a reviewed clock-target policy is added. The paper-to-PhenoBench mapping is: - Task: `chronological_age` - Predictor: system/modal participant features - Strategy: supervised regression on canonical PB splits - Secondary artifact: biological-age residual or centered BA score `biological_age_residual_years` is defined as `predicted_chronological_age_years - chronological_age_years`; positive values mean older-than-chronological. This is a row-level derived output, not a separate observed target. `biological_age_centered_residual_years` is defined as `predicted_chronological_age_years - mean(predicted_chronological_age_years | split, sex, floor_chronological_age_year)`. This mirrors the paper's sex and 1-year age-bin centering shape, but uses split-local centering for PB artifacts so train predictions do not set the eval-row baseline. **What a row here does not support.** The target is a date of birth. Predicting it well means a modality carries age-correlated structure - nothing more. In particular the derived residual is **not a validated measure of biological age or of ageing rate**: a positive residual says a model placed a participant older than their years on the information it was given, which is a property of the model and the modality as much as of the participant. Treating it as a clinical ageing phenotype needs a reviewed clock-target policy, longitudinal validation against outcomes, and a stability analysis that none of these rows provide. ## Related HPP Measurements Demonstrated in HPP and persisted in this repository, from an internal review record, test split: - V1 sleep reaches test R2 = 0.273 (95% CI [0.237, 0.305]) on a cohort of 9,453. - V1 Nightingale reaches R2 = 0.124 (95% CI [0.093, 0.152]) on 10,128. - V1 CGM reaches R2 = 0.094 (95% CI [0.064, 0.125]) on 11,483. - The four-modality fusion (CGM + sleep + diet + …) reaches R2 = 0.421 (95% CI [0.365, 0.467]) on 5,491. **Read the fusion number against its cohort, not against the others.** It scores 5,491 participants where the single-modality rows score 9,453-11,483, because requiring every modality shrinks the eligible population by roughly half. A fusion arm can therefore beat a single modality partly by scoring an easier or simply different set of people; the paired comparison on matched support is what separates the two explanations, and it is not in these numbers. **This card's snapshot carries no paired deltas against a demographic floor**, unlike every other card in this set - the tracks are single-modality with no floor comparator recorded, which is expected here because age is the target and cannot be a covariate. So the R2 values above are raw predictability, and there is no committed margin to quote. Proposed rather than shown: that any of this reflects biological ageing rather than modality-specific age structure. Nothing committed distinguishes them. **These are `reviewable` rows, not `final` ones.** Measured across the committed leaderboards: all 787 rows carry `curation_state: reviewable` and none is `final`. A complete fresh v0 re-run is committed to, so every figure in this section is pre-re-run evidence that the re-run supersedes - persisted and citable today, not a frozen result. **Which curation states the Figure 1 export accepts is the export track's live question, not this card's**; read the two counts above, not a gate. ## Baselines And Ceilings - **Floor**: no-modality-feature baseline. `NoFeaturesPredictor` contributes no modality features, and `DemographicFloorStrategy` fits only allowed demographic covariates: sex and BMI here, because age is the target and is forbidden as a covariate. For OOF modality comparisons, use named support cohorts so the floor and modality row score the same participants. - **Sleep track**: `HppModalityFeaturePredictor x RidgeCVStrategy` over loader-native WatchPAT sleep features after conservative age-leak exclusions. - **CGM track**: `HppModalityFeaturePredictor x RidgeCVStrategy` over loader-native CGM summary features after conservative age-leak exclusions. - **Nightingale track**: `HppParticipantFeaturePredictor x RidgeCVStrategy` over an explicit NMR lipid-biomarker allowlist. This is not the paper's 3,098-cluster LC-MS lipidomics feature space. - **Serum lipidomics OOF track**: `HppModalityFeaturePredictor x SexStratifiedLightGBMCVStrategy` over annotated HPP serum lipidomics features. This is a closer blood-lipid analogue than Nightingale NMR, but still not the paper's gated 3,098-cluster WIS LC-MS matrix. - **Renal OOF track**: `HppParticipantFeaturePredictor x SexStratifiedLightGBMCVStrategy` over long-form blood-test pivots for creatinine, urea, sodium, and potassium. This is the first Reicher/Bar age-clock panel in PB that exercises explicit `blood_tests:` feature specs. - **Frailty OOF track**: `HppParticipantFeaturePredictor x SexStratifiedLightGBMCVStrategy` over hand-grip, DXA lean-mass, height, and weight fields. This maps the paper's frailty-system route, but the current PB config exposes raw fields rather than derived best-grip, ALM, or ALMI features. - **Olink proteomics OOF track**: `OlinkProteomicsPredictor x SexStratifiedLightGBMCVStrategy` over the whole Olink Reveal NPX proteome (~1,000 proteins after a >=0.10 detection-rate filter), the plasma-proteome analogue of the serum-lipidomics OOF track. Below-LOD cells are left NaN under the default `nan_below_lod` policy and train-median imputed in-strategy. A ridge baseline (`OlinkProteomicsPredictor x SexStratifiedRidgeCVStrategy`) pairs the synthetic smoke. Config: `chronological_age_olink_proteomics_lightgbm_oof.operational source.example.yaml`. - **Sex-stratified age-clock tracks**: `Hpp*FeaturePredictor x SexStratifiedRidgeCVStrategy` variants for sleep, CGM, and Nightingale. These fit one model per sex stratum and use BMI as the remaining demographic covariate. - **LightGBM age-clock tracks**: `Hpp*FeaturePredictor x SexStratifiedLightGBMCVStrategy` variants for sleep, CGM, and Nightingale. These are closer to Reicher/Bar 2024: one LightGBM regressor per sex stratum and train-split OOF predictions for BA artifacts. Validation/test rows are still PB split rows predicted by full train-stratum models, not full paper OOF across the complete matched cohort. - **LightGBM OOF tracks**: `chronological_age_*_lightgbm_oof` configs score every row in the predictor-support cohort via sex-stratified K-fold predictions. They are the closest current PB analogue to the paper's OOF reporting protocol, while still using PB loader feature routes and caveats. - **Reviewed artifact track**: `ParticipantFeatureArtifactPredictor x SexStratifiedLightGBMCVStrategy` for regenerated Reicher-style per-system matrices. For paper-parity probes, use a named matched-support cohort, sex-only demographic features for stratification, `subsample_freq=1`, and inner early stopping controls. - **MMFM/fusion embedding tracks**: `ParticipantEmbeddingPredictor` or `FusionTuplePredictor` over imported Adva/Alon MMFM artifacts. These configs use `split_source: artifact` and `cohort_spec: predictor_support`, because the artifact defines the available participant set and train/validation split. In this context, "fusion" means imported MMFM/fusion embeddings, not naive concatenation of HPP feature tables. The current starter artifacts are CGM+sleep static handoff artifacts; a diet+activity+CGM+sleep artifact should use the same predictor contract once available. - **Fusion lift diagnostic**: `FusionTuplePredictor x RidgeCVDeltaR2Strategy` logs demographics-only R2, demographics+embedding R2, delta-R2, bootstrap CI, and permutation p-value as scalar metrics. - **Olink proteome lift diagnostic**: `OlinkProteomicsPredictor x RidgeCVDeltaR2Strategy` logs demographics-only R2, demographics+proteome R2, delta-R2, bootstrap CI, and permutation p-value on one predictor-support cohort/fold set. ## Caveats - Do not include `age`, age-matched percentiles, z-scores, baseline trackers, visit counters, birth fields, or model-derived BA scores in predictor features. - PB canonical participant splits are not the paper's sex-stratified 3-fold OOF scheme. Leaderboard rows are paper-inspired unless a paper-faithful strategy is added. - Reicher/Bar 2024 trained sex-specific LightGBM models. The sex-stratified RidgeCV starter rows match the sex-specific model shape but use RidgeCV. LightGBM rows require the optional `boosting` extra. OOF LightGBM rows should be compared separately from canonical validation/test leaderboard rows. - `SexStratifiedLightGBMCVStrategy` supports optional `subsample_freq`, `early_stopping_validation_fraction`, and `early_stopping_rounds` controls for closer reproduction of paper-style LightGBM age-clock training. Defaults preserve the simpler PB starter behavior. - Biological age is not an observed target here. Treat BA residuals/scores as derived outputs until a reviewed imported-clock target contract exists. - `biological_age_centered_residual_years` is still a derived output, not an observed target. It is split-local PB centering, not the paper's full sex-stratified OOF BA workflow. - Exact blood-lipid parity is gated by the paper's LC-MS lipid-cluster route. - Diagnosis and menopause association scorecards belong to follow-up analysis artifacts, not the primary scalar age leaderboard row. - Olink proteomics enters as a whole-proteome feature route (`olink_proteomics`, information source `olink`). Organ/system panel clocks are deferred; any future route needs separate review and must not become a `panel` kwarg on the whole-proteome predictor. - Olink selection bias: participants with QC-passed Olink skew slightly older and narrower in age (mean ~51.8 vs ~50.4 years for non-Olink), so the proteomic clock describes the assayed sub-population, not the full HPP cohort. The narrower age spread can also depress absolute R2. - Olink NPX is log2 and relative; below-LOD (~13.5% of QC-passed cells) is left-censored, not missing-at-random. The default `nan_below_lod` policy drops those cells to NaN for leakage-safe train-median imputation; `keep` and `half_lod_impute` are ablation arms that change the surviving feature count and can inflate R2 via below-LOD noise. - For the Olink lift diagnostic, `r2_demo_embed` is a pooled (non-stratified) Ridge and is NOT the sex-stratified LightGBM OOF headline R2. Do not cross-compare the two; the delta-R2 is only valid internally against its own demographics-only floor. ## Metrics Primary metric is R2 for chronological-age prediction. Secondary MAE in years is essential for clinical readability. For aging-clock work, persist and audit `biological_age_residual_years` and `biological_age_centered_residual_years` when available, but do not treat them as observed targets. ## References - Reicher, N. et al. 2024, "Phenome-wide associations of human aging uncover sex-specific dynamics". - Reicher et al. (2024), phenome-wide associations of human ageing, *Nature Aging*, . Local copy in **HPP documentation**: `assets/papers/hpp/reicher_2024_phenome_wide_associations_of_human_aging/`. - HPP documentation replication package: `research_runs/paper-replicate/reicher_2024_phenome_wide_associations_of_human_aging/`. - Argentieri, M.A. et al. (2024), proteomic aging clock predicting mortality and age-related disease (UK Biobank Olink), . ## v0 Final-Test Status - Four approved representation-probe Benchmark Tracks use V1 Nightingale, CGM, sleep, or CGM + sleep + diet + activity fusion, with sex and BMI but no age input. - Sex-stratified Ridge and LightGBM produced eight remotely verified rows. Best held-out R² values were 0.124, 0.094, 0.273, and 0.421, respectively. - The fusion artifact is checksum- and four-source-provenance-bound, but this release does not independently audit every upstream pretraining variable for age-correlated side channels. Treat 0.421 as a representation-probe result. - Full contract and results: `05-sleep-gut-aging.md`. ## Olink Proteomic Clock — Initial Validation First real-data validation of the Olink proteome route (baseline `00_00_visit`, `min_detection_rate=0.10`, `nan_below_lod`, `cohort_spec=predictor_support`). These are validation evidence, not published leaderboard rows. The implementation and validation discussion are recorded in PR #149; no dedicated leaderboard rows are committed for these numbers. **The figures below are an uncommitted validation read and are not reproducible from this repository.** No leaderboard row, review JSON or report under `reports/` holds them, so a reader cannot re-derive them from the tree and a re-run will not reproduce them without the PR's configuration. A pull-request discussion is a record of a conversation, not a persisted derivation - the distinction this repository has already paid for once, when a widely quoted overlap figure was removed for having no derivation behind it. They are kept because the result is real and worth knowing; they are marked so nobody quotes them as committed evidence. See Open Questions. - **Whole-proteome OOF clock** (`sex_stratified_lightgbm_cv`, `eval_split=oof`): R² = 0.726, cohort n = 4,870 (train/validation/test 3,376 / 740 / 754). - **Proteome vs demographic floor** (`ridge_cv_delta_r2`, same cohort/folds, 1000-bootstrap / 1000-permutation): demographics-only R² = 0.017, demographics+proteome (pooled Ridge) R² = 0.792, ΔR² = 0.774 (95% CI [0.742, 0.807], permutation p < 0.001). The proteome supplies essentially the entire age signal over the sex+BMI floor (~2% → ~79% variance explained). The pooled-Ridge 0.792 is not the OOF headline 0.726 — different model and eval split; do not cross-compare. ## Clinically Meaningful Variants **No categorical variant of chronological age is meaningful, and that is not the interesting question here - the derived residual is.** Age is a continuous known quantity; binning it would create an age-group classification Task whose labels are already available, which measures nothing. The live variant question is whether `biological_age_residual_years` should become a target in its own right, and the answer is not yet: - **It is a model output, not a measurement.** The residual depends on which modality the clock saw and how well it fits, so two clocks disagree about the same participant by construction. A "biological age" Task would be scoring one model against another model's output. - **Thresholds on it have no clinical standing.** There is no validated cutoff for an accelerated-ageing label, and any percentile bin would be relative to this cohort's own clock. - **The centering convention is load-bearing.** `biological_age_centered_ residual_years` centres within split, sex and 1-year age bin, so the value a participant receives depends on who else is in their split - which is defensible for an artifact and unacceptable for a target. The card already records the right disposition: log these as derived Strategy outputs or artifact metadata, and gate a clock target behind a reviewed policy. ## Open Questions - Which imported biological-age clock target, if any, should become a separate reviewed task? - Should OOF age-clock rows be separated from standard validation/test leaderboard rows in the public UI? - Which age-leak feature families need automated checks beyond the current card warnings? - **Should the Olink validation figures be persisted or dropped?** They are the only numbers on this card with no derivation in the tree. Persisting them under `reports/` or as a review JSON would make them citable; leaving them qualified, as now, keeps a real result visible at the cost of it being unverifiable here. Alon's call on 2026-08-15 was to qualify rather than strip or persist, because persisting is a run's worth of work outside the evidence track. - Does the four-modality fusion beat the single-modality rows **on matched support**? Its cohort is 5,491 against 9,453-11,483, so the committed comparison cannot separate a better model from an easier population.