Predicting post-stroke functional outcome using explainable machine learning and integrated data

Area of Science:

• Neurology
• Biomarkers
• Machine Learning

Background:

• Functional outcomes after acute ischemic stroke (AIS) vary significantly.
• Current prognostic scores may not fully capture complex patient factors.

Purpose of the Study:

• To predict 3-month functional outcome after AIS using machine learning.
• To identify key clinical characteristics and blood biomarkers influencing stroke prognosis.
• To leverage explainable AI for understanding model performance drivers.

Main Methods:

• Trained machine learning models (XGBoost, MLP, L1/L2 logistic regression) on data from 506 AIS patients.
• Utilized Shapley additive global explanations for feature importance assessment.
• Compared model performance using AUROC and AUPRC metrics.

Main Results:

• All models demonstrated high predictive accuracy (AUROC 0.900-0.906).
• The MLP model showed superior precision-recall performance and sensitivity.
• Stroke severity (NIHSS) was the primary predictor, with BD-tau and inflammation markers providing additional prognostic information.

Conclusions:

• Machine learning accurately predicts functional outcomes post-AIS.
• Blood biomarkers, including BD-tau and inflammation proteins, enhance prognostic capabilities beyond clinical stroke severity.
• These findings support the potential for integrating biomarkers into individualized AIS prognostication strategies.