SHAP and Beyond
summary
A gradient-boosting model was developed with Amsterdam UMC to predict ICU readmission risk (AUC 0.82) using thousands of features from vital signs, labs and clinical time series, and embedded in Pacmed software to support discharge decisions. To make the model usable by clinicians the team engineers clinically meaningful, interpretable features, uses SHAP to explain individual predictions, and builds a front-end interpretability layer; they ran user studies with ICU doctors to identify misunderstandings and iterate on explanations before wider validation and implementation.
AI-written summary · about
chapters (14)
- 0:00Healthcare AI mission
- 1:22ICU discharge challenges
- 3:08Readmission risk prediction
- 4:35SHAP and model transparency
- 7:50Model transparency
- 11:20Time series feature engineering
- 14:37Patient risk explanation methods
- 17:45Game theory feature contribution
- 20:54User testing with doctors
- 24:02Complexity versus explainability
- 26:31Prediction uncertainty
- 29:01Uncertain patient discharges
- 30:45Value of explainable AI
- 31:27Conclusion
related talks
23mTransferring clinical prediction models across hospitals...
33mBringing life and motion to AI Explainability
29mBart Steverink & Laurens Schinkelshoek
35mResponsible implementation of Machine Learning models for life-and-death decision making
31mBeyond the Black Box: Interpreting ML Models with SHAP
35mSHAPtivating Insights: unravelling blackbox AI models
54mCrossing the Last Mile with Predictable Hospital Readmissions
35mTaking a peek under the hood - Interpreting black box model