Scholarlink

<p>Abnormal birth weight, including macrosomia and low birth weight, constitutes a significant global health burden associated with both immediate neonatal risks and long-term metabolic complications. However, accurate prediction remains challenging due to fragmented clinical indicators and a lack of model interpretability, limitations often inherent in traditional statistical approaches that struggle with complex, high-dimensional data. This study developed interpretable machine learning models that integrate multifaceted maternal and fetal characteristics to enhance prediction accuracy and facilitate causal analysis. Key features were identified through statistical significance and correlation tests, and causal inference was conducted using G-computation. Among 14 evaluated models, XGBoost achieved optimal performance with AUC values of 0.997 for macrosomia and 0.992 for low birth weight. Feature importance analysis revealed distinct pathogenic pathways—metabolic and biometric factors were most predictive for macrosomia, whereas placental and hemodynamic factors dominated low birth weight prediction. These insights provide a robust and interpretable framework for early risk detection, supporting personalized antenatal management and improved perinatal healthcare strategies.</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

Predicting macrosomia and low birth weight with interpretable machine learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月2日

查看全文
  • Min Cui,
  • Haiying Yang,
  • Bingxin Wang,
  • Jin Zhang,
  • Qianqian Chu,
  • Chunxiao Zhang,
  • Zhaomin Yao,
  • Li Wang

摘要

Abnormal birth weight, including macrosomia and low birth weight, constitutes a significant global health burden associated with both immediate neonatal risks and long-term metabolic complications. However, accurate prediction remains challenging due to fragmented clinical indicators and a lack of model interpretability, limitations often inherent in traditional statistical approaches that struggle with complex, high-dimensional data. This study developed interpretable machine learning models that integrate multifaceted maternal and fetal characteristics to enhance prediction accuracy and facilitate causal analysis. Key features were identified through statistical significance and correlation tests, and causal inference was conducted using G-computation. Among 14 evaluated models, XGBoost achieved optimal performance with AUC values of 0.997 for macrosomia and 0.992 for low birth weight. Feature importance analysis revealed distinct pathogenic pathways—metabolic and biometric factors were most predictive for macrosomia, whereas placental and hemodynamic factors dominated low birth weight prediction. These insights provide a robust and interpretable framework for early risk detection, supporting personalized antenatal management and improved perinatal healthcare strategies.