Random forest-based prediction of stroke outcome.

Area of Science:

• Neurology
• Biomedical Informatics
• Data Science in Healthcare

Background:

• Stroke outcome prediction is crucial for patient management.
• Machine learning offers novel approaches to analyze complex clinical data.
• Existing models may not fully capture the nuances of different stroke types.

Purpose of the Study:

• To develop a machine learning model for predicting 3-month mortality and morbidity in stroke patients.
• To identify key clinical, biochemical, and neuroimaging predictors of stroke outcomes.
• To compare model performance across ischemic stroke (IS) and intracerebral hemorrhage (ICH) subtypes.

Main Methods:

• Prospective data collection from 6022 stroke patients (IS and ICH) at a European tertiary hospital.
• Utilized Random Forest (RF) machine learning algorithm to build predictive models.
• Identified key variables including NIHSS scores and axillary temperature for outcome prediction.

Main Results:

• RF models demonstrated high accuracy in predicting mortality for IS+ICH (AUC 0.904) and IS (AUC 0.909) groups.
• The ICH group showed lower prediction accuracy for mortality (AUC 0.7104).
• Morbidity prediction showed no significant difference between IS and IS+ICH groups, with statistically significant differences noted between IS and ICH groups.

Conclusions:

• Machine learning, specifically RF, is effective for predicting long-term mortality and morbidity in stroke patients.
• Key predictors include NIHSS scores and admission temperature.
• Model performance varies across different stroke subtypes, highlighting the need for tailored approaches.