Scholarlink

<p>Predicting clinical outcomes, such as mortality, is critical for decision-making in Intensive Care Units (ICUs). However, the development of accurate machine learning (ML) models is hampered by data scarcity and class imbalance inherent in this domain. This paper presents and validates a methodology to overcome these limitations by generating and selecting high-fidelity tabular synthetic data. The methodology consists of four stages: (1) proactive data quality assurance at the source, (2) class balancing employing oversampling, (3) large-scale synthetic data generation, and (4) final selection of the optimal dataset. The selection process is guided by a unified fidelity index that integrates both statistical similarity to the original data and predictive utility for the classification task, ensuring that the generated data are realistic and functionally effective. The effectiveness of the methodology was empirically validated using ten real-world, scarce, and imbalanced datasets from an ICU. The results demonstrate that classification models trained with the selected synthetic data achieved significant and consistent improvements across all key performance metrics, including Area Under Curve (AUC-ROC), F1 score and Accuracy, compared to models trained solely on the original data. This paper concludes that the proposed methodology provides a practical and effective way to generate data that allows the building of robust and reliable clinical predictive models in critical care settings with limited data.</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

Methodology for Generating High Fidelity Synthetic Data to Improve Mortality Prediction for Patients in Intensive Care Units

  • 期刊论文

  • 发表日期：

    2026年4月18日

查看全文
  • Marcos Díaz Bastida,
  • Ramiro A. Pérez Vázquez,
  • Rafael Bello Pérez

摘要

Predicting clinical outcomes, such as mortality, is critical for decision-making in Intensive Care Units (ICUs). However, the development of accurate machine learning (ML) models is hampered by data scarcity and class imbalance inherent in this domain. This paper presents and validates a methodology to overcome these limitations by generating and selecting high-fidelity tabular synthetic data. The methodology consists of four stages: (1) proactive data quality assurance at the source, (2) class balancing employing oversampling, (3) large-scale synthetic data generation, and (4) final selection of the optimal dataset. The selection process is guided by a unified fidelity index that integrates both statistical similarity to the original data and predictive utility for the classification task, ensuring that the generated data are realistic and functionally effective. The effectiveness of the methodology was empirically validated using ten real-world, scarce, and imbalanced datasets from an ICU. The results demonstrate that classification models trained with the selected synthetic data achieved significant and consistent improvements across all key performance metrics, including Area Under Curve (AUC-ROC), F1 score and Accuracy, compared to models trained solely on the original data. This paper concludes that the proposed methodology provides a practical and effective way to generate data that allows the building of robust and reliable clinical predictive models in critical care settings with limited data.