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

Area of Science:

• Neurology
• Health Services Research
• Medical Informatics

Background:

• Thrombolysis is a critical treatment for acute ischemic stroke.
• Significant variations in thrombolysis use exist between hospitals.
• Understanding these variations is key to improving stroke care equity.

Purpose of the Study:

• To investigate the factors contributing to between-hospital differences in thrombolysis administration for emergency stroke admissions in England and Wales.
• To quantify the impact of patient characteristics versus hospital-level factors on thrombolysis utilization.

Main Methods:

• Utilized machine learning (XGBoost) on the Sentinel Stroke National Audit Programme (SSNAP) dataset (n=88,928 patients, 2016-2018).
• Employed SHAP (SHapley Additive exPlanations) values for model interpretability.
• Analyzed patient features and hospital identity as predictors of thrombolysis receipt.

Main Results:

• Thrombolysis use varied widely (7%-49%) between hospitals.
• Key factors influencing thrombolysis included arrival-to-scan time, stroke severity, onset time accuracy, pre-stroke disability, and patient age.
• The majority of between-hospital variance was attributed to hospital-level factors, not patient population differences.

Conclusions:

• Explainable machine learning identified in-hospital processes and clinical judgment variations as primary drivers of thrombolysis use disparities.
• Improving standardized in-hospital stroke care pathways and decision-making protocols is crucial.
• Addressing these systemic factors can help reduce geographical inequities in acute stroke treatment.