An improved stacking model for predicting myocardial infarction risk in imbalanced data

Area of Science:

• Cardiology
• Machine Learning
• Data Science

Background:

• Early diagnosis and treatment of myocardial infarction (MI) are crucial for reducing disease severity.
• Imbalanced datasets in disease prediction commonly lead to suboptimal outcomes with conventional models.
• Developing robust MI risk prediction models for imbalanced data presents a significant challenge.

Purpose of the Study:

• To introduce a novel stacked model, 2GDNN-FL-Stacked, designed to predict MI risk effectively using imbalanced datasets.
• To mitigate the impact of data imbalance through random under-sampling and cost-sensitive learning.
• To enhance prediction accuracy by integrating multiple machine learning algorithms.

Main Methods:

• A novel stacked model (2GDNN-FL-Stacked) was developed, combining 2GDNN-FL, CatBoost, RandomForest, and LightGBM.
• Data imbalance was addressed using random under-sampling and cost-sensitive techniques.
• Ablation experiments were conducted to validate the contribution of each model component.

Main Results:

• The 2GDNN-FL-Stacked model demonstrated significant improvements in key performance metrics.
• Matthews Correlation Coefficient (MCC) increased by 15.70%, F1-score by 9.81%, and Area Under the ROC Curve (AUC) by 8.11% compared to baseline models.
• Ablation studies confirmed the effectiveness and necessity of all integrated components for optimal performance.

Conclusions:

• The 2GDNN-FL-Stacked model effectively addresses the challenge of MI risk prediction in imbalanced datasets.
• The proposed method offers superior performance over single models and provides valuable insights for clinical decision-making.
• This approach shows potential for aiding clinicians in making more informed decisions regarding heart attack risk assessment.