Algo-Model

MetaboGuard Results and Terminology Guide

Current status

MetaboGuard has a valid self-supervised representation/deviation artifact, but does not currently have a validated future cancer or diabetes risk model.

The current artifact may be used to:

• generate latent metabolic representations;
• calculate deviation percentiles;
• identify measurements contributing to reconstruction deviation;
• benchmark unsupervised tabular representations.

It must not be described as predicting future development from current NHANES data.

Self-supervised artifact

Item	Result
Unlabelled training rows	50,000
Raw features	25
Latent dimensions	16
Validation reconstruction loss	0.0565

Post-hoc cross-sectional association checks:

Label	AUROC	AUPRC	Interpretation
Any-cancer prevalence	0.699	0.169	Representation association only
Type 2 diabetes proxy	0.923	0.675	Current-state association only
Type 1 proxy	0.909	0.135	Unvalidated research proxy

These values are not development-risk performance.

Pancreatic model invalidation

An audit against the official NHANES codebook found that the first pooled dataset used MCQ230 code 39 as pancreatic cancer. The official coding is:

• 29 = Pancreas (pancreatic)
• 39 = Other

The former metrics and MetaboGuard-XGB v1 artifact are invalid and must not be quoted, benchmarked, published or used for inference.

Corrected cohort

The corrected nhanes_multicycle_v2.csv contains:

Definition	Positive cases
Pancreatic cancer across the pooled population	19
Pancreatic cancer among participants with diabetes	7
Usable diabetic positives after required-field cleaning	6
Pancreatic cancer diagnosed 0-3 years after diabetes	2

These event counts are too small for train/test splitting, calibration, temporal validation or model comparison.

Safety behaviour

MetaboGuard now:

1. Excludes nhanes_multicycle.csv and nhanes_merged.csv from API discovery.
2. Uses the versioned corrected datasets ending in _v2.csv.
3. Requires at least 20 usable positive and 20 usable negative cases.
4. Stores a SHA-256 dataset signature in every new artifact.
5. Retrains or refuses inference when the dataset signature changes.
6. Blocks public model export when the corrected cohort is underpowered.

Priority A variables now available

Variable	MetaboGuard field	Coverage in pooled cohort
Smoking status	smoking_status	59,745
Current smoker	current_smoker	59,745
Alcohol status	alcohol_status	51,467
Average drinks/day	average_drinks_per_day	34,282
Haemoglobin	CBC_LBXHGB	87,554
Platelet count	CBC_LBXPLTSI	87,552
ALT	BIOPRO_LBXSATSI	67,442
Alkaline phosphatase	BIOPRO_LBXSAPSI	67,536
Creatinine	BIOPRO_LBXSCR	67,542
Reciprocal HbA1c term	hba1c_reciprocal_100	68,644
Squared HbA1c term	hba1c_squared	68,644

These variables improve the feature schema but cannot compensate for only six usable positive diabetic pancreatic-cancer cases.

Is higher better?

Output	Meaning	Direction
AUROC	Ranking discrimination	Higher is better; 0.5 is random
AUPRC	Precision-recall performance	Higher is better; compare with prevalence
AUPRC lift	AUPRC divided by prevalence	Higher is better; 1× is baseline
Brier score	Squared probability error	Lower is better
Feature importance	Model dependence on a feature	Larger means more influence, not causality

No corrected pancreatic-risk values are currently reported for these metrics.

What is still valid

• TCGA-CDR prognosis experiments remain separate and are unaffected by the NHANES site-code correction.
• NHANES diabetes and any-cancer descriptive analyses remain available.
• The corrected v2 datasets and variable-coverage audit are valid.
• The paper-derived genetic, protein and metabolite candidate catalogues remain valid research inputs.

Required next dataset

The brief requires a cohort containing:

• a sufficiently large incident pancreatic-cancer endpoint;
• exact diabetes and cancer diagnosis dates;
• at least 20 positive cases for exploratory fitting, preferably far more;
• repeated HbA1c, glucose, BMI and weight measurements;
• smoking and routine blood tests;
• ideally genotype, CA19-9 and/or the prioritised serum proteins.

UK Biobank or an equivalent linked clinical/genetic cohort is required for the primary MetaboGuard model. NHANES should remain a feature-engineering and population-context dataset.

How to read each output (2026-08-04 terminology contract)

Output	What it is	What it is not	Safe sentence
metabolic_deviation_score	Distance of a profile from the training reference (0.7 × reconstruction + 0.3 × latent distance, both robust-scaled)	A probability, a severity grade, or a diagnosis	“This profile is unusual relative to our reference sample.”
reference_percentile	Rank of that score inside the training reference distribution	A population percentile (survey weights are not applied)	“It sits above the 95th percentile of our training reference.”
latent_representation	16 learned numbers summarising the inputs	Interpretable biology; individual dimensions have no assigned meaning	“It is the encoding we use for downstream research.”
top_deviation_features	Features contributing most to reconstruction error	Causes, or clinically abnormal values	“Most of the unusualness comes from HbA1c and triglycerides.”
cross_sectional_association_probability	Probability that someone with this profile already has a recorded diagnosis in NHANES	A future-risk probability; a screening result	“Profiles like this are more common among people who already report a diagnosis.”
Post-hoc AUROC / AUPRC	Separation of prevalent cases on the holdout partition	Prevention or early-warning performance	“The representation carries cross-sectional signal.”
Baseline reconstruction MSE	How well an unsupervised method models the feature distribution	Disease prediction skill	“The encoder models the data slightly better than a matched PCA.”
Flag Jaccard between methods	Overlap of the top-5 % most-unusual profiles	Agreement about disease	“Which profiles look unusual depends on the method, so clinicians adjudicate.”

Deprecated: cancer_risk_probability is retained for one release as an alias of cross_sectional_association_probability. Do not quote it; the name implies future risk the data cannot support.

Reading the research-pass outputs (added 2026-08-04)

Feature eligibility tiers (data observation)

Tier	What it means for interpretation
usable_now	Present, plausible, ≥50% coverage among adults, no blocking drift or availability gap.
qualified_use	Usable with the stated caveat printed alongside any result: low coverage, cycle availability gap, cycle-level drift, implausible-value burden, or a declared evidence gap.
unavailable	Absent or effectively unmeasured. Any statement about it is unsupported.
prohibited	Outcome label, label-derived column, or post-diagnosis TCGA context. Never an input.

Current file: 17 usable_now, 8 qualified_use (C-peptide, glucose, insulin, hs-CRP, triglycerides, LDL, HOMA-IR, race/ethnicity), 0 unavailable, 16 prohibited.

Cluster outputs (model association)

Output	What it is	What it is not
cluster_id	A phenotype identifier such as cluster_2	A cancer type, cancer site, disease subtype or severity grade
prototype_median_profile	Median raw feature values inside the phenotype	A patient, a reference range or a target
top_distinguishing_panel	Features whose robust standardised difference from the reference is largest	Causes, or evidence that the panel detects disease
membership_confidence	GMM posterior, or k-means relative margin	A probability of disease
posthoc_label_summary	Prevalence of an already-recorded diagnosis per phenotype, suppressed below 50 cases	Future risk, incidence, or validation of the phenotype
status: no_stable_clusters	A result: nothing passed the gates	A bug, or a reason to lower the thresholds

If the status is no_stable_clusters, the correct sentence is: “No stable metabolic phenotype survived our stability and negative-control checks on this data.” Do not describe candidate solutions from the table as findings.

Negative controls

A gating control above 0.30 means the clustering largely reproduces that nuisance variable. Age and sex associations are expected to be non-zero in metabolic data; only values above the threshold mark a solution as a data artefact. The exact assay-availability pattern is shown for transparency and never gates a decision.

Evidence rows (published evidence)

Read three fields together before quoting any number: stage_or_lead_time, validation_status and limitations. A high AUC measured at diagnosis says nothing about early detection; the lead-time column is where the honest answer lives. Rows without a source URL or with an ungraded evidence_grade are research-only and must not reach a clinician-facing surface. See EVIDENCE_AND_CLAIMS.md for the allowed and denied statement lists.

Gated off entirely: any 1/3/5-year horizon output. /api/v1/prevention-future-risk returns HTTP 409 with the event-count gate report, and prevention-capabilities.longitudinal_heads_enabled is false.

Professor assessment payload format (2026-08-05)

Professor-facing API/UI surfaces now render a shared envelope so interpretation is explicit and consistent:

• current_profile_assessment: current-state meaning (deviation band or cross-sectional association framing).
• standout_factors: feature-level contributions for this profile.
• data_readiness: missing inputs with priority and rationale.
• research_association: prevalence/proxy-only status text.
• safety_contract: diagnostic_status=non_diagnostic, future_risk=disabled.

Example reading patterns

1. Within reference range: “Current profile assessment shows within-reference-range deviation. No model warning; still not evidence that disease is absent.”
2. Elevated deviation with missing key data: “Deviation is elevated, and data readiness lists missing high-impact fields before deeper interpretation.”
3. High deviation with risk-disabled behavior: “Profile is highly unusual, but future-risk and cancer-type outputs remain disabled on cross-sectional capability state.”

   Future-risk rows (simulation only)

The dashboard’s future-risk panel shows one row per horizon (1, 3 and 5 years). Every row is one of exactly two things:

1. A simulated calibrated estimate — cumulative incidence from a simulation-only artefact, shown alongside the raw uncalibrated value and the selected model name. It is synthetic. It is not a patient’s risk, not a diagnosis and not evidence of early detection.
2. An explicit abstention — the horizon did not pass the 50-event / 50-non-event gate, so no number is produced. Abstention is a result, not a failure to load.

Reading the numbers:

• Calibrated vs raw: calibration is fitted on a synthetic validation split. Absolute values are not population-calibrated when the cohort used declared enrichment strata.
• Competing outcomes: incidence is cause-specific with death treated as a competing event, so the target-outcome curve does not absorb deaths.
• Uncertainty: bootstrap confidence intervals, calibration slope/intercept, Brier, average precision, subgroup metrics and negative controls live in the run artefact’s results.json — not per individual synthetic history.
• Real patients: the clinical future-risk endpoint returns HTTP 409. The panel states this.

The dashboard’s future-risk panel shows one row per horizon (1, 3 and 5 years). Every row is one of exactly two things:

1. A simulated calibrated estimate — cumulative incidence from a simulation-only artefact, shown alongside the raw uncalibrated value and the selected model name. It is synthetic. It is not a patient’s risk, not a diagnosis and not evidence of early detection.
2. An explicit abstention — the horizon did not pass the 50-event / 50-non-event gate, so no number is produced. Abstention is a result, not a failure to load.

Reading the numbers:

• Calibrated vs raw: calibration is fitted on a synthetic validation split. Absolute values are not population-calibrated when the cohort used declared enrichment strata.
• Competing outcomes: incidence is cause-specific with death treated as a competing event, so the target-outcome curve does not absorb deaths.
• Uncertainty: bootstrap confidence intervals, calibration slope/intercept, Brier, average precision, subgroup metrics and negative controls live in the run artefact’s results.json — not per individual synthetic history.
• Real patients: the clinical future-risk endpoint returns HTTP 409. The panel states this.

The dashboard’s future-risk panel shows one row per horizon (1, 3 and 5 years). Every row is one of exactly two things:

1. A simulated calibrated estimate — cumulative incidence from a simulation-only artefact, shown alongside the raw uncalibrated value and the selected model name. It is synthetic. It is not a patient’s risk, not a diagnosis and not evidence of early detection.
2. An explicit abstention — the horizon did not pass the 50-event / 50-non-event gate, so no number is produced. Abstention is a result, not a failure to load.

Reading the numbers:

• Calibrated vs raw: calibration is fitted on a synthetic validation split. Absolute values are not population-calibrated when the cohort used declared enrichment strata.
• Competing outcomes: incidence is cause-specific with death treated as a competing event, so the target-outcome curve does not absorb deaths.
• Uncertainty: bootstrap confidence intervals, calibration slope/intercept, Brier, average precision, subgroup metrics and negative controls live in the run artefact’s results.json — not per individual synthetic history.
• Real patients: the clinical future-risk endpoint returns HTTP 409. The panel states this.

This site is open source. Improve this page.