Predicting short-term outcomes in atrial-fibrillation-related stroke using machine learning

Area of Science:

• Computational neurology
• Cardiovascular research
• Medical informatics

Background:

• Atrial fibrillation (AF) is linked to adverse outcomes in stroke patients.
• Accurate prognostic prediction is crucial for early management of AF-related strokes.

Purpose of the Study:

• To develop and validate machine learning models for predicting short-term outcomes in AF-related stroke.
• To identify key prognostic factors contributing to stroke outcomes in AF patients.

Main Methods:

• Utilized two independent datasets (K-ATTENTION and KUSR) for internal and external validation.
• Developed and compared a logistic regression model with tree-based and multi-layer perceptron (MLP) machine learning models.
• Employed the Area Under the Receiver Operating Characteristic Curve (AUROC) for performance evaluation and Shapley Additive Explanation (SHAP) for variable importance.

Main Results:

• Machine learning models, particularly MLP, significantly outperformed logistic regression in predicting 3-month unfavorable functional status (AUROC 0.890 internal, 0.859 external).
• Both ML models showed superior prediction for 3-month mortality in internal validation compared to logistic regression.
• The initial National Institute of Health and Stroke Scale score was the most significant predictor for both unfavorable outcomes and mortality.

Conclusions:

• Explainable machine learning models offer reliable prediction of short-term outcomes for AF-related stroke.
• These models effectively identify high-risk patients, aiding in targeted interventions and improved stroke care.