A Stroke Risk Detection: Improving Hybrid Feature Selection Method.

Area of Science:

• Biomedical Informatics
• Machine Learning for Healthcare
• Cardiovascular Disease Research

Background:

• Stroke is a leading cause of mortality, and early risk detection is crucial but complicated by numerous risk factors.
• Existing stroke risk detection methods face limitations in effective feature selection, hindering accurate prediction.
• Identifying critical risk factors is essential for developing effective and concise predictive models.

Purpose of the Study:

• To address limitations in feature selection for stroke risk detection.
• To propose and evaluate a novel feature selection method, weighting- and ranking-based hybrid feature selection (WRHFS), for identifying ischemic stroke risk factors.
• To enhance the accuracy and efficiency of stroke risk prediction models.

Main Methods:

• Developed the weighting- and ranking-based hybrid feature selection (WRHFS) method, integrating filter algorithms within a wrapper approach.
• Utilized filter-based feature selection models (e.g., standard deviation, Pearson correlation, Fisher score, information gain, Relief, chi-square) as candidates.
• Evaluated WRHFS performance using sensitivity, specificity, accuracy, and Youden index on a dataset of 792 patient samples with 28 initial features.

Main Results:

• The WRHFS method successfully selected 9 important features from the original 28, significantly outperforming baseline methods.
• The selected 9 features demonstrated a cumulative contribution of 0.51.
• WRHFS achieved high performance metrics: 82.7% sensitivity, 80.4% specificity, 81.5% accuracy, and a 0.63 Youden index.

Conclusions:

• A novel feature selection method (WRHFS) was developed and validated for stroke risk detection.
• The WRHFS method enables the creation of effective and parsimonious models by identifying the most significant risk factors.
• Findings offer practical implications for improving stroke risk assessment and patient care.