This page summarizes what the tool is based on, who it applies to, where the data came from, how the model was built, and where interpretation should stop.
Why this tool exists.
The study focuses on people with diagnosed diabetes who still have fasting blood glucose below 7.0 mmol/L, a group that can be missed by threshold-based screening alone.
The main hypothesis was that risk in this range is not linear, and that residual organ injury may shift from glucocentric mechanisms toward hemodynamic and behavioral drivers.
A key study finding was a physiological fasting glucose window of 4.86-5.24 mmol/L rather than a simple “normal equals safe” interpretation.
Development and external verification cohorts.
Model development used a multi-province Chinese cohort coordinated by Zhujiang Hospital, with internal geographic splitting across provinces.
External verification used NHANES 2013-2020 to test transportability and support survey-weighted mechanistic analyses.
The manuscript reports a low-prevalence DNFG identification setting, so calibration quality matters as much as discrimination.
High-level modeling approach used in the manuscript.
Primary learner: histogram-based gradient boosting decision tree ensemble.
Class imbalance handling: partitioned negatives plus single-pass SMOTE inside training folds.
Probability calibration: isotonic regression to make predicted probabilities more clinically interpretable.
Interpretability: SHAP for feature ranking and nonlinear dependence patterns.
Study-level metrics from the manuscript.
External validation ROC-AUC: 0.859.
PR-AUC: 0.306, approximately 10 times the baseline prevalence of 0.03.
The intended use is screening and stratification in a low-prevalence scenario, not stand-alone diagnosis.
Main findings translated into operational statements.
FBG Safety Window
FBG showed a U-shaped risk relationship within the sub-diagnostic range. The lowest-risk window was identified at 4.86-5.24 mmol/L.
Hemodynamic Takeover
SBP acted as a shared mediator of cross-organ burden, mediating 30.3% of liver fibrosis risk and 20.1% of kidney function decline in the reported mediation analyses.
Behavioral Signal
Smoking exposure and lower protective lifestyle signals aligned with higher-risk phenotypes and were treated as modifiable contributors rather than background variables.
Diagnosed diabetes with fasting blood glucose below 7.0 mmol/L.
Users seeking screening, stratification, or follow-up context.
Interpretation where SBP, smoking, exercise, and FBG context can be reviewed together.
Do not use this tool as a replacement for diabetes diagnosis, treatment initiation, or medication changes.
Do not treat the output as causal proof of kidney or liver injury without formal clinical evaluation.
Interpret cautiously when key inputs are missing or when the user falls outside the study-like population.
Current implementation boundary in this project.
This site currently displays calibrated risk output, study-aligned interpretation, and transparency information.
It does not yet expose individual SHAP values, formal LCA class assignment, or full intervention re-simulation from the manuscript workflow.