Automatic Feature Selection for Imbalanced Echocardiogram Data Using Event-Based Self-Similarity.

Area of Science:

• Cardiology
• Biomedical Engineering
• Data Science

Background:

• Imbalanced echocardiogram datasets pose challenges for cardiovascular disease (CVD) prediction, leading to biased machine learning models.
• Optimal feature selection and robust classification are crucial for enhancing CVD prognosis accuracy.
• Echocardiogram data, visual sound wave signals, and patient treatment data were utilized.

Purpose of the Study:

• To introduce an event-based self-similarity approach for enhanced automatic feature selection in imbalanced echocardiogram data.
• To identify critical features correlated with CVD progression using self-similarity patterns.
• To improve the accuracy and reduce bias in machine learning models for CVD prognosis.

Main Methods:

• Echocardiogram data was classified into nine categories.
• Recursive Feature Elimination (RFE) was employed for feature selection.
• XGBoost, CATBoost, and a Random Forest (RF)-voting ensemble classifier were trained and evaluated.

Main Results:

• XGBoost and CATBoost models achieved accuracies of 84.3% and 88.4%, respectively.
• A voting ensemble model improved feature selection and predictive accuracy.
• Age, aorta (AO), left ventricle (LV), and left atrium (LA) were identified as critical prognostic features.

Conclusions:

• Feature selection techniques are vital for handling imbalanced datasets and reducing bias in automated prognosis.
• Machine learning-driven echocardiogram analysis shows potential for enhancing patient care through accurate, data-driven assessments.