🧠 How Machine Learning Predicts Stroke Risk

About this simulation

Gradient-boosted models such as XGBoost don't learn a single formula — they learn an ensemble of shallow decision trees, each one correcting the errors of the trees before it. Every patient's record is pushed through all the trees, their small votes are summed, and the total is squeezed through a sigmoid curve to give a probability between 0% and 100%.

This simulation uses a simplified, teaching-only version of that idea — a smooth logistic surface standing in for the boosted ensemble — so you can see how the shape of the risk landscape shifts as clinical inputs change. It is not a diagnostic tool.

What you're looking at

• Risk landscape — the terrain's height at every point is the model's predicted risk for a patient of that age (one axis) and glucose level (the other), holding blood pressure and lifestyle factors fixed at your slider settings.
• Patient marker — the glowing sphere sits exactly on the surface at your current patient's age and glucose, colour-coded from blue (low risk) to red (high risk).
• Contribution bars — each bar shows how much one feature is currently pushing the log-odds score up (red, toward stroke) or down (blue, protective), the same way SHAP-style explainers break a model's prediction into per-feature contributions.
• Tree ensemble — the small trees behind the scene represent the boosted ensemble; the number lit red approximates how many of them are "voting" toward elevated risk.
• Risk dial — the mounted display reads the combined probability and NHS-style risk band (Low / Moderate / High).

Controls

• Age — patient age in years (30–90).
• Glucose — fasting glucose in mg/dL (70–300).
• Systolic BP — systolic blood pressure in mmHg (90–200).
• Currently smokes / Diagnosed hypertension — binary lifestyle and clinical flags, exactly the kind of categorical features a boosted-tree model splits on.
• ↺ Reset patient — returns all inputs to a moderate-risk baseline. Drag to rotate, scroll to zoom.

Did you know?

Real clinical stroke-risk models trained on UK datasets typically combine dozens of features — age, blood pressure, glucose, BMI, smoking status, atrial fibrillation and more — and are validated against calibration curves before any NHS-style screening tool would use their output to flag a patient for review.