Project overview

What KERI is building.

The project explores whether combinations of metabolic, clinical, and behavioral variables can reveal early warning patterns linked to cancer among people with diabetes.

Research aim

Identify whether multi-variable metabolic patterns can support earlier risk stratification for cancer in diabetes contexts.

Method stack

The current workflow combines causal reasoning, a predictive baseline, and data preparation around the NHANES merged dataset.

Clinical value

The project is framed around earlier identification of higher-risk individuals so the signal can be reviewed before symptoms become advanced.

Variables of interest

Signals to examine in the data

Diabetes timing

Type, onset, and duration of diabetes.

Metabolic markers

HbA1c, insulin, C-peptide, and related biomarkers.

Patient profile

Age, sex, obesity, weight change, and broader clinical context.

Documentation hub

Research docs, translated for readers

This page includes a maintained summary of the core research documents so the Org Site remains readable even when the docs folder is not available in this repository.

Each block below mirrors the current project documentation and is updated to reflect what is valid, what is exploratory, and what still requires a larger longitudinal dataset.

Methodology log

Active

MetaboGuard uses a hybrid self-supervised workflow: a denoising autoencoder learns latent patient representations, then post-hoc association heads test whether those representations carry useful signal.

  • Encoder shape: 25 raw features, 55 transformed dimensions, 16 latent dimensions.
  • Training data: 50,000 unlabeled adult NHANES rows.
  • Current approved capability: cross-sectional representation and deviation scoring only.
Open methodology source

Results and status guide

Guarded

The self-supervised artifact is valid for deviation and representation benchmarking, but the corrected pancreatic endpoint remains underpowered for a deployable risk model.

  • Any-cancer cross-sectional check: AUROC 0.699, AUPRC 0.169.
  • Type 2 proxy check: AUROC 0.923, AUPRC 0.675.
  • Corrected diabetic pancreatic positives after required-field cleaning: 6 cases.
Open results guide source

Prevention model specification

Specification

The prevention specification defines what the system can and cannot claim, with strict dataset-governance gates before any disease-risk release is allowed.

  • Warning output is non-diagnostic and clinician-reviewed.
  • Leakage-sensitive variables are explicitly deny-listed.
  • Future risk heads require longitudinal outcomes and event-count thresholds.
Open model spec source

Evidence brief

Literature-backed

The project evidence brief maps key design choices to published epidemiology and biomarker literature, and flags where evidence is still incomplete.

  • New-onset diabetes has materially higher pancreatic risk than long-standing diabetes.
  • Weight-loss and glycemic dynamics are high-value candidate signals.
  • NHANES and TCGA-CDR are complementary, but do not create patient-level trajectories.
Open evidence brief source

Variable catalogue

Roadmap

The variable catalogue translates recent papers into implementation priorities so readers can track exactly which predictors are available now versus what needs richer cohorts.

  • Priority A variables now integrated include smoking, alcohol, CBC, ALT and creatinine.
  • Nonlinear HbA1c behavior is explicitly modeled as a top research focus.
  • Genetic and omics candidates are separated into higher-complexity future tiers.
Open variable catalogue source

Reader guide

What the outputs mean today

Metric meanings

Metric Meaning Direction
AUROC Ranking discrimination across thresholds. Higher is better.
AUPRC Precision-recall performance under class imbalance. Higher is better; compare with prevalence.
Brier score Squared probability error. Lower is better.

Terminology in plain language

  • Cross-sectional: one-time measurements, useful for associations.
  • Longitudinal: repeated follow-up over time, required for future-risk claims.
  • Deviation score: how unusual a profile is versus the training reference population.
  • Leakage: giving a model information that directly encodes the outcome.
  • Underpowered cohort: too few positive cases to support stable validation.

Evidence of progress

  • Invalidated old endpoint mapping and corrected pancreatic code construction.
  • Moved pipeline to corrected versioned datasets ending in _v2.csv.
  • Added dataset signatures and artifact safety gates to block stale model reuse.
  • Expanded predictor coverage with Priority A clinical variables from literature review.
  • Documented explicit release policy for non-diagnostic versus future-risk artifacts.

Open Source Development

Code and model development

Code repository

The project is open source and available on GitHub. The repository includes scripts for data preparation, model training, and evaluation.

View on GitHub Open dashboard

Model development

The model is being developed in Python using libraries such as scikit-learn, pandas, and NumPy. The workflow includes data cleaning, feature engineering, and model evaluation.

The project aims to provide a baseline predictive model that can accurately identify individuals at higher risk of cancer based on their metabolic and clinical profiles.