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.
| 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).
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.
| 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 |
In feature-importance order shown in the paper:
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).
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 |
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.
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.
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.
The 2026 study identified 39 proteins with consistent direction in both the clinical case-control and model-risk comparisons.
XPNPEP2, NCAN, PLA2G7, CRTAC1, ADAMTS8, ITGAV,
KITLG, NTRK3, PLTP, MOG, DKK3, RGMB, GPA33.
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).
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.
Routine, scalable variables:
Clinical v2 plus:
Clinical v2 plus PC, T2DM and/or NODM-PC PRS. Validate by geography or external cohort and include ancestry principal components.
Used only after clinical high-risk triage:
Avoid post-diagnosis leakage from:
Those variables belong to MetaboGuard’s separate prognosis experiments, not the NODM early-detection model.
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.