Risk prediction of 30-day mortality after stroke using machine learning: a nationwide registry-based cohort study.
Wenjuan Wang1, Anthony G Rudd2, Yanzhong Wang2,3,4
1School of Population Health & Environmental Sciences, Faculty of Life Science and Medicine, King's College London, London, UK. wenjuan.wang@kcl.ac.uk.
BMC Neurology
|May 27, 2022
Summary
Machine learning models accurately predict 30-day stroke mortality for risk stratification and quality improvement. The XGBoost model demonstrated superior performance in temporal validation, outperforming traditional logistic regression models.
Area of Science:
- Medical Informatics
- Machine Learning in Healthcare
- Stroke Medicine
Background:
- Developing accurate machine learning (ML) models for 30-day stroke mortality is crucial for risk stratification.
- These models can serve as benchmarks for improving stroke care quality.
Purpose of the Study:
- To develop and validate ML models for predicting 30-day stroke mortality.
- To assess their utility in risk stratification and as quality improvement benchmarks.
Main Methods:
- Utilized UK Sentinel Stroke National Audit Program data (2013-2019).
- Developed XGBoost and Logistic Regression (LR) models with 30 variables, validated temporally.
- Compared performance using discrimination, calibration, reclassification, Brier scores, and Decision-curves.
Main Results:
- The XGBoost model achieved the highest AUC (0.895) and lowest Brier score (0.069) in temporal validation.
- XGBoost outperformed the LR reference model and LR with elastic net.
- Models showed good calibration, with XGBoost effectively reclassifying low-risk patients.
Conclusions:
- Models using 30 variables are valuable for stroke care quality improvement benchmarking.
- Machine learning models offer a slight performance advantage over traditional methods.
Keywords:
30-day mortalityMachine learningOutcomesQuality improvementRisk predictionStatistical analysisStroke
