Optimizing Machine Learning-Based Classification of Cardiac Arrhythmias Through Feature Selection

Introduction: The clinical complexity of cardiac arrhythmias drives the adoption of Machine Learning (ML) for diagnosis. However, model efficacy critically depends on identifying the most predictive features. This study investigates advanced feature selection methods to isolate optimal parameters, aiming to enhance the accuracy and efficiency of arrhythmia classification models. Using optimal feature selection, this study identifies key Electrocardiogram (ECG) and clinical predictors to enhance ML model accuracy in detecting cardiac arrhythmias. Material and Methods: This computational study employed a two-phase feature refinement using Correlation Feature Selection (CFS) with Best-First Search, distilling best 50 features and 27 elite predictive features from global and localized ECG characteristics. Multiple machine learning models were then developed and assessed based on this optimized feature set. Results: The results demonstrated significant improvements in classification accuracy with feature selection. Random forest achieved an accuracy of 69.46% without feature selection, which increased to 73.67% with the top 50 features and further improved to 75.66% with the elite features. Similarly, LogitBoost showed a remarkable increase in accuracy from 74.55% to 80.97% when using the elite features. Conclusion: Considering the increase in cardiac diseases and their treatment costs, finding the most important features and using Artificial Intelligence (AI) will improve the screening and diagnosis of these patients. Also, electronic health record data and the design of medical decision support systems can be helpful in helping to treat and improve patient management.