Scholarlink

<p>Machine Learning (ML) techniques assist cardiologists in computationally analyzing cardio-health data. As the ML models are resource-intensive, it is often desirable to achieve acceptable results without major compromise in accuracy with smaller volumes of data. The challenge lies in the identification of the significant features out of the comprehensive feature set. This study delves into the complexities of feature explanation methodologies in ML, utilizing the advanced techniques of SHAP, SHAPASH, and LIME. SHAP provides stable importance scores for each feature, whereas SHAPASH streamlines and illustrates these explanations for easier understanding. LIME provides localized understanding of how particular attributes influence single predictions. Collectively, these tools boost confidence in model predictions, enabling the recognition of the most important features in cardiac arrest classification. By combining these techniques, the research identifies ten of the most contributing features for predicting cardiac arrest have been shortlisted amongst the total feature vector, which encompasses 53 features. To achieve this objective, a thorough assessment is carried out on some key ML models, namely, XGBoost, SVM, Random Forest (RF), and Logistic Regression (LR). Upon careful examination, it becomes evident that XGBoost stands out as the most powerful performer in all aspects, demonstrating exceptional skill, especially in handling the difficulties presented by imbalanced datasets. Significantly, its accuracy in forecasting the dominant class (“Discharge”) exceeds that of its peers. In this study, we also explore ensemble modeling by employing both “Voting” and “Stacking” approaches. The fusion of XGBoost with SVM and RF demonstrates the ultimate combination of ensemble synergy, thereby achieving a remarkable accuracy of 89.32% in the voting mechanism and 90.33% in stacking. The adoption of the Synthetic Minority Oversampling Technique (SMOTE) is crucial in addressing the inherent imbalance in the dataset.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CardiacXAI: Explainable Ensemble AI Schema for Unveiling Critical Features in Cardiac Arrest Classification

  • 期刊论文

  • 发表日期：

    2025年9月9日

查看全文
  • Soma Mitra,
  • Mauparna Nandan,
  • Samarjit Roy,
  • Debashis De

摘要

Machine Learning (ML) techniques assist cardiologists in computationally analyzing cardio-health data. As the ML models are resource-intensive, it is often desirable to achieve acceptable results without major compromise in accuracy with smaller volumes of data. The challenge lies in the identification of the significant features out of the comprehensive feature set. This study delves into the complexities of feature explanation methodologies in ML, utilizing the advanced techniques of SHAP, SHAPASH, and LIME. SHAP provides stable importance scores for each feature, whereas SHAPASH streamlines and illustrates these explanations for easier understanding. LIME provides localized understanding of how particular attributes influence single predictions. Collectively, these tools boost confidence in model predictions, enabling the recognition of the most important features in cardiac arrest classification. By combining these techniques, the research identifies ten of the most contributing features for predicting cardiac arrest have been shortlisted amongst the total feature vector, which encompasses 53 features. To achieve this objective, a thorough assessment is carried out on some key ML models, namely, XGBoost, SVM, Random Forest (RF), and Logistic Regression (LR). Upon careful examination, it becomes evident that XGBoost stands out as the most powerful performer in all aspects, demonstrating exceptional skill, especially in handling the difficulties presented by imbalanced datasets. Significantly, its accuracy in forecasting the dominant class (“Discharge”) exceeds that of its peers. In this study, we also explore ensemble modeling by employing both “Voting” and “Stacking” approaches. The fusion of XGBoost with SVM and RF demonstrates the ultimate combination of ensemble synergy, thereby achieving a remarkable accuracy of 89.32% in the voting mechanism and 90.33% in stacking. The adoption of the Synthetic Minority Oversampling Technique (SMOTE) is crucial in addressing the inherent imbalance in the dataset.