Scholarlink

To reduce problems during pregnancy and improve outcomes for newborns, it is very important to accurately and quickly assess the health of the fetus. Cardiotocography (CTG) is a non-invasive way to keep an eye on the health of a fetus. However, manual interpretation is still open to bias, differences between observers, and mistakes in diagnosis, especially when working with large and unbalanced datasets. This study introduces a correlation-based, L2-regularized Deep Neural Network (DNN) framework for multi-class Fetal Health Risk Prediction (L2R4FHRP), classifying CTG records into Normal, Suspect, and Pathological categories. Pearson’s correlation coefficient ( \(|r| > 0.2\) ) is used to keep only statistically significant predictors. This reduces multicollinearity, eliminates unnecessary features, and accelerates calculations. The preprocessed dataset contains 2,126 occurrences and 21 numerical characteristics generated from CTG. It is standardized and encoded using one-hot encoding to represent labels. To mitigate overfitting and ensure the model’s applicability to diverse datasets, the proposed architecture employs fully connected layers with ReLU activation, dropout regularization, and L2 weight decay ( \(\lambda = 0.001\) ). The Adam optimizer (learning rate of \(\eta = 0.001\) ) and categorical cross-entropy loss is employed for training to improve the network. This aligns the expected and actual class distributions as closely as possible. Experimental testing indicates that the correlated feature subset surpasses the whole feature set in outcome prediction, with an accuracy of 96.45% compared to 95.51%, while constantly improving precision, recall, and F1-score. The analysis of the confusion matrix reveals that the suggested technique facilitates the distinction between the “Suspect” and “Pathological” classes, highlighting its therapeutic importance. The optimal method to equilibrate model complexity, interpretability, and generalization performance is to use correlation-based feature selection in conjunction with L2-regularized deep learning.

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

L2-Regularized Deep Neural Network Model for Robust Multi-class Fetal Health Risk Prediction

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Kanak Kumar,
  • Anshul Verma,
  • Pradeepika Verma

摘要

To reduce problems during pregnancy and improve outcomes for newborns, it is very important to accurately and quickly assess the health of the fetus. Cardiotocography (CTG) is a non-invasive way to keep an eye on the health of a fetus. However, manual interpretation is still open to bias, differences between observers, and mistakes in diagnosis, especially when working with large and unbalanced datasets. This study introduces a correlation-based, L2-regularized Deep Neural Network (DNN) framework for multi-class Fetal Health Risk Prediction (L2R4FHRP), classifying CTG records into Normal, Suspect, and Pathological categories. Pearson’s correlation coefficient ( \(|r| > 0.2\) ) is used to keep only statistically significant predictors. This reduces multicollinearity, eliminates unnecessary features, and accelerates calculations. The preprocessed dataset contains 2,126 occurrences and 21 numerical characteristics generated from CTG. It is standardized and encoded using one-hot encoding to represent labels. To mitigate overfitting and ensure the model’s applicability to diverse datasets, the proposed architecture employs fully connected layers with ReLU activation, dropout regularization, and L2 weight decay ( \(\lambda = 0.001\) ). The Adam optimizer (learning rate of \(\eta = 0.001\) ) and categorical cross-entropy loss is employed for training to improve the network. This aligns the expected and actual class distributions as closely as possible. Experimental testing indicates that the correlated feature subset surpasses the whole feature set in outcome prediction, with an accuracy of 96.45% compared to 95.51%, while constantly improving precision, recall, and F1-score. The analysis of the confusion matrix reveals that the suggested technique facilitates the distinction between the “Suspect” and “Pathological” classes, highlighting its therapeutic importance. The optimal method to equilibrate model complexity, interpretability, and generalization performance is to use correlation-based feature selection in conjunction with L2-regularized deep learning.