A stroke-risk model can output a single number — "23% risk" — but that number alone gives a clinician nothing to check, question, or explain to a patient. SHAP (SHapley Additive exPlanations) values, borrowed from cooperative game theory, split that prediction into a sum of feature contributions: how much did age push the score up, how much did blood pressure push it up, how much did a healthy BMI push it down — all starting from an average "baseline" risk across the population.
|SHAP value| — bigger bars are bigger drivers of the decision.SHAP values are named after Lloyd Shapley, who won the 2012 Nobel Memorial Prize in Economic Sciences for Shapley values in cooperative game theory — originally designed to fairly split a payout among players, now repurposed to fairly split a prediction among features.
A UK stroke-risk model's prediction is decomposed into a 3D SHAP waterfall: a baseline population risk, followed by stacked bars showing exactly how much each clinical feature pushed the score up or down to reach the final prediction.
SHAP values split a single opaque risk score into additive, per-feature contributions rooted in Shapley game theory. Red bars raise predicted risk above the population baseline; purple bars lower it — and they always sum exactly to the final prediction.
Pick a preset patient or drag the age, blood pressure, glucose and BMI sliders yourself, toggle smoking status, and watch every bar resize and recolour live. Click any bar to read its exact contribution.
Because SHAP values are additive and consistent, a clinician can check whether a model's reasoning matches known medicine — e.g. that hypertension should raise risk — before ever trusting its number.