Methods
SHAP (Shapley additive explanations)
Also called: SHAP values, SurvSHAP, explainable AI, XAI
SHAP explains an individual prediction by splitting it into contributions from each input feature, based on Shapley values from game theory.
SHAP (SHapley Additive exPlanations) assigns each input feature a contribution to one specific prediction. Contributions add up to the difference between that prediction and the model's average output, so a clinician can see which factors pushed this patient's risk up and which pulled it down.
The values come from Shapley values in cooperative game theory, which fairly divide a payout among players. Exact versions exist for tree-based models; model-agnostic versions such as Kernel SHAP approximate them for any model. Time-dependent extensions such as SurvSHAP(t) show how contributions change across the follow-up period of a survival model.
SHAP explains the model, not the biology. A feature with a large contribution is predictive in the data, which is not the same as being a cause that would change the outcome if altered.
Tools that use this
- Craniopharyngioma Outcomes StudioPredict tumor recurrence and permanent diabetes insipidus after craniopharyngioma surgery, with SHAP explanations and an extent-of-resection trade-off.
- Aneurysm Recanalization StudioIndividualized risk of major recanalization after endovascular coiling over 12 to 48 months, with packing density what-ifs. C-index 0.83 on 3,530 aneurysms.
- Adjacent Segment Disease Prediction StudioTime-to-event risk of adjacent segment disease after lumbar fusion at 1, 2, 3 and 5 years, from sagittal alignment, levels fused and surgical technique.
- Open Spina Bifida Frailty and SurvivalFrom findings at birth, predict frailty and the chance of survival at ages 5, 25, 35 and 40 for children born with open spina bifida.
- MatLevel-CDSCounty-level map and decision support for ACOG levels of maternal care: drive time to Level III care, maternity care deserts and postpartum complication risk.
- ManuscriptAIPredict the chance a bioRxiv or medRxiv preprint is published within 90 days to 2 years, with time to publication, journal recommendations and SHAP feedback.
Related terms
Educational reference. Not medical advice and not for clinical decision-making.