A Bayesian Network Model for Predicting Post-stroke Outcomes With Available Risk Factors.

Area of Science:

• Computational biology
• Medical informatics
• Machine learning in healthcare

Background:

• Bayesian networks offer interpretable predictions crucial for medical research.
• Accurate prediction of post-stroke outcomes is essential for patient management.
• Existing methods may lack interpretability or predictive accuracy for complex stroke data.

Purpose of the Study:

• To construct an interpretable and accurate Bayesian network-based inference engine for post-stroke outcomes.
• To forecast functional independence at 3 months and mortality at 1 year after acute stroke.
• To optimize the prediction system using feature selection and rigorous evaluation.

Main Methods:

• Utilized Bayesian network classifiers trained on data from 3,605 acute stroke patients.
• Applied feature selection methods (wrapper-type and information gain) to reduce 76 risk variables.
• Employed hill-climbing search and maximum description length for network structure and parameter optimization.
• Evaluated system performance using the area under the receiver operating characteristic curve (AUC) and sensitivity.

Main Results:

• Bayesian network with wrapper-selected features predicted 3-month functional independence (AUC 0.889, 19 variables) and 1-year mortality (AUC 0.893, 24 variables).
• Bayesian network with information gain-filtered features predicted 3-month functional independence (AUC 0.875) and 1-year mortality (AUC 0.895) using 50 features.
• Achieved high AUC values while maintaining acceptable sensitivity for imbalanced data.

Conclusions:

• The developed Bayesian network system provides accurate and interpretable predictions for post-stroke outcomes.
• Feature selection significantly enhances prediction accuracy and reduces model complexity.
• An online prediction service (Yonsei Stroke Outcome Inference System) was established to support clinical application.