Algo-Model

MetaboGuard Column Dictionary

This dictionary covers identifiers, outcomes, survey-design fields and every feature used by the public clinical-only model. The machine-readable equivalent is model_artifacts/huggingface/metaboguard-risk-xgboost/feature_schema.json.

Outcomes and identifiers

Column	Meaning	Values / units	Is higher better?
global_participant_id	Stable pooled identifier	survey_cycle:SEQN	Not applicable
SEQN	NHANES participant identifier inside a cycle	Integer	Not applicable
survey_cycle	NHANES collection period	Text, e.g. 2017-March2020	Not applicable
survey_cycle_index	Ordered cycle number	0-9	Not a health measure
PancreaticCancer	Corrected target derived from MCQ230A-D code 29	1 positive, 0 negative	Not applicable
NODM_PancreaticCancer	Pancreatic cancer diagnosed 0-3 years after diabetes	1 positive, 0 diabetic control, missing when timing unknown	Not applicable
pancreatic_cancer_diagnosis_age	Age when pancreatic cancer was diagnosed	Years; MCQ240T where available	Not applicable
pancreatic_cancer_minus_diabetes_years	Cancer diagnosis age minus diabetes diagnosis age	Years	0-3 defines NODM-PC
Cancer	Self-reported history of any cancer	1 yes, 0 no	Not applicable
Diabetes	Self-reported diagnosed diabetes	1 yes, 0 no	Not applicable

Survey design

Column	Meaning	Use
DEMO_SDMVPSU	Masked variance pseudo-primary sampling unit	Survey variance estimation
DEMO_SDMVSTRA	Masked variance stratum	Survey variance estimation
combined_mec_weight_1999_2020	Combined examination weight	Descriptive examined-sample population estimates only
DEMO_WTMEC2YR	Original two-year examination weight	Cycle-specific analyses
DEMO_WTMEC4YR	NCHS 1999-2002 bridge weight	Required for the first two cycles
DEMO_WTMECPRP	2017-March 2020 pre-pandemic weight	Pre-pandemic cycle

The combined MEC weight is not a fasting-subsample weight and is not used to train the prediction model.

Demographics and body measurements

Column	Meaning	Values / units	Direction
DEMO_RIDAGEYR	Age	Years	Higher age was associated with higher risk; not inherently “worse”
DEMO_RIAGENDR	NHANES sex code	1 male, 2 female	Categorical; higher is meaningless
DEMO_RIDRETH3	Race/ethnicity category	NHANES code	Categorical; higher is meaningless
BMX_BMXBMI	Body mass index	kg/m²	Higher generally means more adiposity; relationship may be nonlinear
BMX_BMXWAIST	Waist circumference	cm	Higher generally means greater central adiposity
Obesity	Derived BMI threshold	1 if BMI ≥30	1 indicates obesity

Diabetes history

Column	Meaning	Values / units	Direction
DIQ_DID040	Reported age at diabetes diagnosis	Years	Lower means earlier onset
DIQ_DIQ160	Prediabetes questionnaire response	NHANES code	Categorical
DIQ_DIQ170	Diabetes-risk questionnaire response	NHANES code	Categorical
DIQ_DIQ180	Diabetes-risk questionnaire response	NHANES code	Categorical
diabetes_duration_years	Current age minus diagnosis age	Years	Risk is duration-dependent, not simply higher/lower
recent_diabetes_onset	Diabetes duration ≤3 years	1 yes, 0 no	1 is a recognised risk-enrichment signal
new_onset_diabetes	Alias of recent-onset flag	1 yes, 0 no	1 indicates recent onset
diabetes_subtype	Exploratory subtype proxy	0 non-diabetic, 1 Type-1-like, 2 Type-2-like	Categorical; unvalidated heuristic

Glycaemic and metabolic laboratories

Column	Meaning	Values / units	Direction
GHB_LBXGH	Glycated haemoglobin, HbA1c	Percent	Higher means poorer average glycaemic control
GLU_LBXGLU	Fasting plasma glucose	mg/dL	Higher means higher fasting glucose
INS_LBXIN	Fasting insulin	µU/mL	Interpret with glucose; higher is not automatically better/worse
CPEP_LBXCPSI	Fasting C-peptide	SI units	Higher means more endogenous insulin secretion; early cycles only
homa_ir	Glucose × insulin / 405	Derived index	Higher suggests greater insulin resistance
elevated_hba1c	HbA1c ≥6.5%	1 yes, 0 no	1 indicates diabetic-range HbA1c
fasting_hyperglycemia	Fasting glucose ≥126 mg/dL	1 yes, 0 no	1 indicates diabetic-range glucose

Lipids and inflammation

Column	Meaning	Values / units	Direction
TRIGLY_LBXTR	Triglycerides	mg/dL	Higher is generally metabolically adverse
TRIGLY_LBDLDL	LDL cholesterol	mg/dL	Higher is generally metabolically adverse
HDL_LBDHDD	HDL cholesterol	mg/dL	Higher is generally metabolically favourable
TCHOL_LBXTC	Total cholesterol	mg/dL	No universal “better” direction in this model
HSCRP_LBXHSCRP	High-sensitivity C-reactive protein	mg/L	Higher indicates inflammation; pancreatic-risk evidence is weak

Paper-supported Priority A additions

Column	Meaning	Values / units	Direction
smoking_status	Harmonised smoking category	0 never, 1 former, 2 current	Categorical; current smoking was higher risk in the cited study
current_smoker	Current smoking flag	1 yes, 0 no	1 means current smoking
alcohol_status	Harmonised alcohol category	0 never/low, 1 ever, 2 current quantified use	Categorical
average_drinks_per_day	Drinks on drinking days	Drinks/day	Higher means greater intake
CBC_LBXHGB	Haemoglobin concentration	g/dL	No simple pancreatic-risk direction
CBC_LBXPLTSI	Platelet count	10³ cells/µL	No simple pancreatic-risk direction
BIOPRO_LBXSATSI	Alanine aminotransferase, ALT	U/L	Higher may indicate liver injury
BIOPRO_LBXSAPSI	Alkaline phosphatase	U/L	Higher may indicate hepatobiliary or bone processes
BIOPRO_LBXSCR	Serum creatinine	mg/dL	Higher generally indicates lower renal filtration
hba1c_reciprocal_100	100 divided by HbA1c	Derived	Lower corresponds to higher HbA1c
hba1c_squared	HbA1c squared	Derived	Nonlinear sensitivity term

Weight history

Column	Meaning	Formula / values	Direction
weight_loss_1yr_lb	Reported one-year weight change	Weight 1 year ago − current weight	Positive means weight loss
significant_weight_loss_flag	Recent loss of at least 10 lb	1 yes, 0 no	1 indicates substantial recent loss
weight_loss_10yr_lb	Reported ten-year weight change	Weight 10 years ago − current weight	Positive means long-term weight loss

These variables do not distinguish intentional from unintentional weight loss.

Interaction and trajectory-proxy features

Column	Formula	Interpretation
age_bmi_interaction	Age × BMI	Combined age/adiposity pattern
waist_bmi_interaction	Waist × BMI	Combined central/general adiposity
hba1c_age_interaction	HbA1c × age	Whether glycaemic signal differs with age
hba1c_diabetes_duration_interaction	HbA1c × diabetes duration	Whether HbA1c signal differs by duration
hba1c_weight_loss_interaction	HbA1c × one-year weight loss	Combined glycaemia/weight-change pattern
hba1c_cycle_age_sex_z	HbA1c z-score within cycle, age band and sex	Cohort-relative value; excluded from the public clinical-only model

For interaction terms, a larger value is not automatically better or worse. Importance means the combination aided ranking.

TCGA-only columns

Column	Meaning	Model role
tcga_cancer_type	TCGA tumour abbreviation	Prognosis models only
tcga_stage_ordinal	Encoded AJCC stage	Prognosis feature
tcga_grade_ordinal	Encoded histological grade	Prognosis feature
tcga_tumor_status	With tumour vs tumour-free	Prognosis feature
tcga_treatment_response	First-course response encoding	Post-treatment prognosis feature
tcga_followup_days, tcga_event	Overall-survival metadata	Excluded from mortality features to prevent leakage
tcga_pfi_days, tcga_pfi_event	Progression metadata	Excluded from progression features to prevent leakage

Missing values

Model features are converted to numeric values and filled with the training-set median. Each temporal fold learns medians without accessing its held-out cycle. The final public artifact stores its medians in feature_schema.json and model.joblib.

Median imputation does not mean a missing test is clinically normal. Missingness often reflects which NHANES cycle or subsample measured the biomarker.

Model-input allowlist (prevention / deviation work)

Defined once in api/self_supervised.py::PREVENTION_FEATURES and validated by api/data_integrity.py. Only these columns may enter the encoder:

DEMO_RIDAGEYR, DEMO_RIAGENDR, DEMO_RIDRETH3, BMX_BMXBMI, BMX_BMXWAIST, GHB_LBXGH, GLU_LBXGLU, INS_LBXIN, CPEP_LBXCPSI, TRIGLY_LBXTR, TRIGLY_LBDLDL, HDL_LBDHDD, TCHOL_LBXTC, HSCRP_LBXHSCRP, CBC_LBXHGB, CBC_LBXPLTSI, BIOPRO_LBXSATSI, BIOPRO_LBXSAPSI, BIOPRO_LBXSCR, smoking_status, alcohol_status, average_drinks_per_day, weight_loss_1yr_lb, weight_loss_10yr_lb, homa_ir.

Categorical members (DEMO_RIAGENDR, DEMO_RIDRETH3, smoking_status, alcohol_status) are one-hot encoded; numeric members are median-imputed with a missingness indicator and robust-scaled on the 10–90 percentile range. All statistics are fit on the training partition only.

Model-input denylist (never an input)

Column / pattern	Why it is denylisted
Cancer, MCQ_MCQ220, MCQ_MCQ230A–D, MCQ_MCQ240T	Outcome and outcome-source columns
PancreaticCancer, NODM_PancreaticCancer, pancreatic_cancer_diagnosis_age, pancreatic_cancer_minus_diabetes_years, same_year_diabetes_pancreatic_cancer	Label-derived; also invalidated targets
Diabetes, DIQ_DIQ010, diabetes_subtype, new_onset_diabetes	Outcome / label-derived
any tcga_* column	Post-diagnosis context (stage, grade, tumour status, treatment response, follow-up time) that cannot inform prevention scoring

data_integrity.is_denylisted_input() enforces both the explicit list and the tcga_ prefix rule, and select_prevention_features() asserts the resulting feature set is disjoint from it.

Label definitions (as implemented)

Label	Definition	Nature
Cancer	MCQ220 == 1 (ever told had cancer)	Prevalent, self-reported, cross-sectional
PancreaticCancer	any of MCQ230A–D == 29 (Pancreas). Code 39 is “Other” and is never counted	Prevalent; 19 cases in 107,622 rows
Diabetes	DIQ010 == 1	Prevalent, self-reported
diabetes_subtype	1 = research-only Type 1 proxy (young onset + insulin), 2 = Type 2 proxy, 0 = no diabetes	Proxy only: no autoantibodies, no approved genetics, no confirmatory C-peptide criteria

Feature eligibility tiers (generated, 2026-08-04)

api/data_reliability.py assigns every candidate column a tier. Regenerate with python data_reliability.py --output ../model_artifacts/reports/data_reliability.json.

Tier	Columns
usable_now	DEMO_RIDAGEYR, DEMO_RIAGENDR, BMX_BMXBMI, BMX_BMXWAIST, GHB_LBXGH, TRIGLY_LBDLDL*, HDL_LBDHDD, TCHOL_LBXTC, CBC_LBXHGB, CBC_LBXPLTSI, BIOPRO_LBXSATSI, BIOPRO_LBXSAPSI, BIOPRO_LBXSCR, smoking_status, alcohol_status, average_drinks_per_day, weight_loss_1yr_lb, weight_loss_10yr_lb (17 columns in the current file)
qualified_use	CPEP_LBXCPSI, GLU_LBXGLU, INS_LBXIN, HSCRP_LBXHSCRP, TRIGLY_LBXTR, TRIGLY_LBDLDL, homa_ir, DEMO_RIDRETH3
unavailable	none in the current file
prohibited	all outcome labels, label-derived columns and every tcga_* column (16 columns)

* tier membership is regenerated per dataset; the authoritative list is the JSON report.

Reasons attached to the qualified_use columns in the current file:

Column	Caveat
CPEP_LBXCPSI	Fasting subsample; measured in a minority of cycles (cycle availability gap).
HSCRP_LBXHSCRP	Cycle availability gap; catalogued evidence shows no consistent association with pancreatic cancer, so treat as a known-weak feature (Bao et al.).
GLU_LBXGLU, INS_LBXIN, homa_ir	Fasting subsample: coverage below the 50% threshold.
TRIGLY_LBXTR, TRIGLY_LBDLDL	Fasting subsample coverage, plus a declared lipid evidence gap in the catalogue.
DEMO_RIDRETH3	Category definitions change across cycles; not comparable as a continuous level.

Columns with catalogued published evidence

Each mapping points at a row in data/evidence/biomarker_evidence.json; read the row’s stage_or_lead_time and limitations before quoting anything.

Column	Evidence row	One-line takeaway
GHB_LBXGH	ev-hba1c-panc-epic	Association strongest within 2 years of diabetes diagnosis.
CPEP_LBXCPSI	ev-cpeptide-panc-michaud	Nonfasting C-peptide associated; fasting insulin was null.
homa_ir	ev-insulin-resistance-panc-toledo	Review-level association, mortality-based attributable risk.
weight_loss_1yr_lb	ev-weightloss-panc-casecontrol, ev-recent-diabetes-weightloss-pdac-jamaoncol-2020	Strong relative association; absolute 4-year incidence still 0.29%.
BMX_BMXBMI	ev-adiposity-panc-pooled, ev-excess-body-fatness-multisite-nejm-2016	IARC sufficient evidence for risk; no early-detection use; weight-loss causality unestablished.
HSCRP_LBXHSCRP	ev-hscrp-panc-bao-null	Explicit negative result.
CBC_LBXPLTSI	ev-thrombocytosis-multisite-bjgp-2017	Non-specific multi-site risk marker; does not indicate site.
TRIGLY_LBXTR	ev-lipids-panc-gap	Declared evidence gap: model it, do not claim it.

Identifiers, splits and capability columns

Column	Meaning
SEQN	NHANES respondent sequence number, unique within a cycle
global_participant_id	"<cycle>:<SEQN>", the participant key used for grouped splitting; 107,622 unique values, no duplicates
survey_cycle, survey_cycle_index, survey_year_midpoint	Pooling metadata; survey_cycle_index is a repeated-cross-section proxy, not a within-patient trajectory
(absent) event_time_days, event	Required for horizon gating. Their absence is why 1/3/5-year heads are disabled

This site is open source. Improve this page.