Algo-Model

MetaboGuard Paper-Derived Variable Catalogue

Purpose

This catalogue converts two recent new-onset diabetes and pancreatic-cancer studies into an actionable feature roadmap for MetaboGuard:

The papers study pancreatic cancer arising within three years of new-onset type-2 diabetes. This is more specific than MetaboGuard’s current NHANES target, which is self-reported prevalent pancreatic cancer among all labelled diabetics.

Highest-priority findings

  1. New-onset timing must remain central. Both papers define new-onset diabetes using a three-year window.
  2. HbA1c should be modelled nonlinearly. The 2025 study selected a nonlinear HbA1c term, not a simple linear coefficient.
  3. Smoking should be added immediately. It survived the 2025 model selection alongside age at diabetes diagnosis and HbA1c.
  4. Routine blood tests can materially expand MetaboGuard. Haemoglobin, platelet count, ALT, creatinine and alkaline phosphatase were candidate predictors but are not currently loaded into the pooled dataset.
  5. Body composition and function matter. The 2026 CatBoost model used walking pace, hip circumference, arm/trunk fat mass and disability status.
  6. Genetic and omics variables should form separate model tiers. They have very different availability and cost from routine clinical variables.
  7. PLTP, CRTAC1 and ITGAV are the leading confirmatory serum candidates. They were supported by cross-platform analyses and clinical assays in the 2026 study.

Clinical variables from the 2025 genetic model

Candidate variables

Paper variable MetaboGuard mapping Current status Recommendation
Age at T2DM diagnosis DIQ_DID040 Available Keep; one of the final predictors
Sex DEMO_RIAGENDR Available Keep as categorical
BMI BMX_BMXBMI Available Keep; test nonlinear effects
Smoking status smoking_status, current_smoker Implemented in v2 0 never, 1 former, 2 current
Glucose GLU_LBXGLU Available Keep; fasting-subsample limitation
HbA1c GHB_LBXGH Available Keep; add nonlinear/spline sensitivity analysis
Haemoglobin CBC_LBXHGB Implemented in v2 87,554 non-null
Triglycerides TRIGLY_LBXTR Available Keep
ALT BIOPRO_LBXSATSI Implemented in v2 67,442 non-null
Creatinine BIOPRO_LBXSCR Implemented in v2 67,542 non-null
Platelet count CBC_LBXPLTSI Implemented in v2 87,552 non-null
Total cholesterol TCHOL_LBXTC Available Keep
Alkaline phosphatase BIOPRO_LBXSAPSI Implemented in v2 67,536 non-null

The final non-genetic model retained only:

The best genetic model added the combined NODM-PC polygenic risk score. Its internal-external C-index was 0.823, compared with 0.749 for the non-genetic model (Zhou et al., 2025).

Desired variables missing from that study

The authors explicitly wanted but could not use:

MetaboGuard already derives one- and ten-year weight change. It does not have true serial glucose/BMI trajectories or reliable pancreatic-symptom coverage.

Clinical variables from the 2026 machine-learning model

Full 22-variable candidate set

Category Variables
Demographic/lifestyle age, sex, ethnicity, smoking, alcohol
Anthropometric/body composition BMI, hip circumference, whole-body fat mass, left-leg impedance, right-arm fat mass, left-arm fat mass, trunk fat mass
Functional/health status usual walking pace; long-standing illness, disability or infirmity
Haematology haemoglobin; immature reticulocyte fraction
Biochemistry apolipoprotein A, glucose, HbA1c, total cholesterol, HDL-C, direct LDL-C

Final 14-variable CatBoost set

In feature-importance order shown in the paper:

  1. HbA1c
  2. Total cholesterol
  3. Age
  4. Apolipoprotein A
  5. Usual walking pace
  6. Immature reticulocyte fraction
  7. Left-arm fat mass
  8. BMI
  9. HDL cholesterol
  10. Long-standing illness, disability or infirmity
  11. Hip circumference
  12. Trunk fat mass
  13. Alcohol status
  14. Sex

The model achieved validation AUROC 0.844. The same-day diabetes/cancer sensitivity analysis reduced AUROC to 0.801, showing the importance of temporal ordering and reverse-causation controls (Yang et al., 2026).

MetaboGuard implementation priorities

Priority A: add to the pooled NHANES clinical model

These should be addressed first because they are routine and potentially obtainable from NHANES:

Variable Reason Expected work
Smoking status Final predictor in the 2025 model Add SMQ files across cycles
Alcohol status Included in the 2026 final model Add ALQ files across cycles
Haemoglobin Used in both candidate sets Add CBC files and harmonise units
Platelet count 2025 candidate Add CBC files
ALT 2025 candidate Add biochemical-profile files
Creatinine 2025 candidate Add biochemical-profile files
Alkaline phosphatase 2025 candidate Add biochemical-profile files
Nonlinear HbA1c Significant nonlinear association Compare spline/reciprocal terms against tree handling
Cancer/diabetes diagnosis interval Needed for true NODM-PC definition Add age/date-at-cancer-diagnosis fields where available

Priority B: audit NHANES availability

Coverage is expected to be incomplete or cycle-specific:

Waist circumference is a reasonable adiposity proxy but is not equivalent to hip circumference or measured regional fat mass.

Priority C: requires a richer longitudinal/genetic cohort

UK Biobank or a comparable linked health-record/genotype cohort is the natural source. These variables cannot be reconstructed reliably from repeated cross-sectional NHANES cycles.

Genetic variables

The 2025 work constructed:

The 33 SNP identifiers are:

rs13303010, rs4655617, rs41276588, rs76263492, rs9873618, rs9854769, rs3887925, rs9854771, rs72501964, rs56337234, rs2853677, rs2736098, rs35226131, rs2735948, rs401681, rs6878122, rs12539264, rs6971499, rs2737226, rs2862954, rs4929965, rs2237895, rs11602873, rs28884829, rs10844518, rs9543325, rs1475655, rs12912777, rs72802342, rs11870735, rs144239147, rs10404726, rs450960.

The machine-readable catalogue also stores each risk allele, fitted beta, p-value and source PRS. These coefficients were estimated in the study cohort and should not be treated as externally validated MetaboGuard coefficients.

Proteomic candidates

The 2026 study identified 39 proteins with consistent direction in both the clinical case-control and model-risk comparisons.

Upregulated in NODM-PC/high-risk groups

XPNPEP2, NCAN, PLA2G7, CRTAC1, ADAMTS8, ITGAV, KITLG, NTRK3, PLTP, MOG, DKK3, RGMB, GPA33.

Downregulated

AMIGO2, EPO, ICAM1, CD22, SH2D1A, PLXNB2, ATF2, CDH1, NRP1, DEFB4A, SIGLEC10, SEMA7A, ROBO1, LGALS4, FCER2, LIFR, LTA, IL18R1, FCRL1, DKK4, FGF23, TCL1A, ADA2, TNFRSF11A, THBS2, TNFRSF13B.

Prioritise PLTP, CRTAC1 and ITGAV for a confirmatory biomarker panel because the paper additionally validated them across Olink, CPTAC, plasma ELISA and tissue assays (Yang et al., 2026).

Metabolomic candidates

The study found 145 metabolites with consistent direction across clinical and model-defined risk comparisons. They are dominated by:

Five highlighted pathway families were:

The full 145-variable list is retained in the machine-readable research candidate data. These platform-specific NMR features should be evaluated as a separate metabolomics model, not median-imputed into the routine clinical model.

MetaboGuard Clinical v2

Routine, scalable variables:

MetaboGuard Clinical-Body v3

Clinical v2 plus:

MetaboGuard Genetic

Clinical v2 plus PC, T2DM and/or NODM-PC PRS. Validate by geography or external cohort and include ancestry principal components.

MetaboGuard Biomarker Confirm

Used only after clinical high-risk triage:

Variables that must not enter an early-detection model

Avoid post-diagnosis leakage from:

Those variables belong to MetaboGuard’s separate prognosis experiments, not the NODM early-detection model.

Implementation outcome

Smoking, alcohol, CBC, routine biochemistry and nonlinear HbA1c terms are now implemented in nhanes_multicycle_v2.csv.

The same audit corrected the pancreatic-cancer site code from 39 (Other) to 29 (Pancreas). This reduced the cohort to 19 pancreatic-cancer cases overall, 7 among diabetics and only 2 meeting the exact three-year NODM-PC definition. Consequently, the next priority is no longer feature expansion: it is acquiring a sufficiently large linked incident pancreatic-cancer cohort.