Skip to main content

Ctrl+K

SAMueL - Stroke Audit Machine Learning

What would other emergency stroke teams do? Using explainable machine learning to understand variation in thrombolysis practice

Introductory content

Glossary of technical terms
Data
Probability, odds, and SHAP values (log odds shifts)
A simple worked example of Shap
Examining interactions between features with SHAP interactions: Using Titanic survival as an example

Predicting thrombolysis use and explaining the predictions with Shap

Optimisation of XGBoost learning rates (regularisation)
XGBoost feature selection
Check correlation between features selected for the model
Measuring accuracy of XGBoost models
Explaining XGBoost model predictions with SHAP values
Fit a single XGBoost model and calculate the SHAP values
Explore the hospital effect on thrombolysis use, using SHAP values from a single XGBoost model
Explore the hospital effect on thrombolysis use, using SHAP main effects from a single XGBoost model
Explore the hospital effect on thrombolysis use, using SHAP values and SHAP main effects from a single XGBoost model
How much of the variation in the hosptials thrombolysis rate can be attributed to the difference in the hospital processes, and to different patient populations?
Calculate and use subset SHAP values (from a single XGBoost model) to attribute hospital IVT variation to patient mix and hospital processes

Using a 10k cohort of patients to compare decision-making

A comparison of 10k cohort thrombolysis rates across hospitals
A comparison of hospital SHAP values and SHAP main effect values for the 10k cohort.
A comparison of expected 10k cohort thrombolysis rates across hospitals: subgroup analysis
A comparison of actual thrombolysis rates across hospitals: subgroup analysis

Index

By Kerry Pearn, Michael Allen, Anna Laws, Richard Everson, and Martin James