Model card
One trained model.
Documented honestly.
This page states exactly what WhyRisk's model is, what data it was trained and measured on, and what it does not do. Figures here come from the project documentation and the held-out evaluation — nothing on this page is estimated or rounded up.
Dataset
CDC BRFSS 2015
Behavioural Risk Factor Surveillance System survey data — 253,680 records, 21 input features per record.
Selected model
LightGBM
Gradient-boosted decision trees, chosen after comparing candidate models on the same data split.
Measured performance
AUC-ROC 0.8307
Measured on a stratified held-out test split (80/20) that the model never trained on. Diabetes prediction only — this figure does not describe any other condition.
Explanations
SHAP
Predictions ship with SHAP feature attributions, so the factors behind a score can be inspected instead of trusted blindly.
Live service status
Checking the ML service…
What is real today
- One trained model: diabetes risk prediction from the 21 BRFSS health indicators.
- SHAP explanations are stored per prediction.
- A separate symptom checker that is a rule-based heuristic — not a trained model, and labelled as such in the product.
Roadmap — not trained models
- Heart disease, stroke, chronic kidney disease, hypertension, asthma and COPD are planned future models.
- They are listed so the scope is transparent: today, WhyRisk predicts diabetes risk only.
How scores are banded
Every score is placed in one of four severity bands — the same bands used everywhere in the app:
Limitations
- The model was trained on United States survey data (BRFSS 2015); performance on other populations has not been validated.
- A score is a screening estimate, not a diagnosis. False positives and false negatives occur.
- WhyRisk is an experimental clinical decision-support and research tool. It is not a certified medical device and does not replace professional diagnosis.