Analyzing the impact of feature selection methods on machine learning algorithms for heart disease prediction

Area of Science:

• Cardiovascular Disease Research
• Machine Learning Applications
• Biomedical Informatics

Background:

• Accurate heart disease prediction is crucial for timely intervention.
• Machine learning models offer potential for improving diagnostic accuracy.
• Optimizing feature selection is key to enhancing machine learning model performance.

Purpose of the Study:

• To evaluate the impact of diverse feature selection techniques on machine learning algorithms for heart disease prediction.
• To identify the most effective feature selection methods and algorithms for this task.
• To compare the performance metrics of various models under different feature selection strategies.

Main Methods:

• Utilized the Cleveland Heart disease dataset.
• Applied sixteen feature selection techniques across filter, wrapper, and evolutionary categories.
• Implemented seven machine learning algorithms: Bayes net, Naïve Bayes, multivariate linear model, Support Vector Machine, logit boost, j48, and Random Forest.
• Evaluated performance using Precision, F-measure, Specificity, Accuracy, Sensitivity, ROC area, and PRC.

Main Results:

• Feature selection yielded performance improvements for some models (e.g., j48) but decreased it for others (e.g., MLP, RF).
• SVM-based filtering methods achieved a best-fit accuracy of 85.5%, with SVM-CFS/information gain/Symmetrical uncertainty showing the highest improvement.
• Filter methods with more selected features excelled in Accuracy, Precision, and F-measures, while wrapper and evolutionary methods boosted Sensitivity and Specificity.

Conclusions:

• Feature selection methods demonstrably influence the efficacy of machine learning models in heart disease prediction.
• Filter methods, especially SVM-based ones, are effective for improving overall accuracy and key classification metrics.
• Wrapper and evolutionary methods offer distinct advantages in enhancing model sensitivity and specificity, suggesting tailored approaches for different predictive goals.