Scholarlink

When left untreated, Gestational Diabetes Mellitus (GDM) can be very dangerous to the health of the mother and child. In this paper, we will present an effective DL classification framework to predict GDM as early as it is possible using clinical data. It consists of three primary steps: preprocessing, feature extraction, and classification. First, Min-Max scaling is employed to normalize the data, bringing features to a common range. Three other feature extraction methods are then used to dimensionally reduce and extract informative patterns: Principal Component Analysis (PCA), Recursive Feature Elimination (RFE), and Autoencoders. To classify, a Long Short-Term Memory (LSTM) neural network is used because it has been noted to be very effective in dealing with sequential and temporal data in medical data. Of the three combinations of pipelines MinMax + PCA + LSTM, MinMax + RFE + LSTM, and MinMax + Autoencoder + LSTM, the MinMax + PCA + LSTM pipeline reports the best results. The performance is measured based on accuracy rate, Area Under the ROC Curve (AUC-ROC), Matthews Correlation Coefficient (MCC), and Cross Entropy Loss (CEL). MinMax + PCA + LSTM model has the best accuracy, AUC-ROC as well as MCC with the least loss, which shows that this model is robust and reliable. The model is verified on the publicly available GDM dataset in Kaggle proving its capabilities to be used in clinical decision support in GDM diagnosis.

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

Deep Learning Approach for Gestational Diabetes Mellitus Classification Using LSTM Neural Network Model

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • V. Venkatesan,
  • K. Rajakumari

摘要

When left untreated, Gestational Diabetes Mellitus (GDM) can be very dangerous to the health of the mother and child. In this paper, we will present an effective DL classification framework to predict GDM as early as it is possible using clinical data. It consists of three primary steps: preprocessing, feature extraction, and classification. First, Min-Max scaling is employed to normalize the data, bringing features to a common range. Three other feature extraction methods are then used to dimensionally reduce and extract informative patterns: Principal Component Analysis (PCA), Recursive Feature Elimination (RFE), and Autoencoders. To classify, a Long Short-Term Memory (LSTM) neural network is used because it has been noted to be very effective in dealing with sequential and temporal data in medical data. Of the three combinations of pipelines MinMax + PCA + LSTM, MinMax + RFE + LSTM, and MinMax + Autoencoder + LSTM, the MinMax + PCA + LSTM pipeline reports the best results. The performance is measured based on accuracy rate, Area Under the ROC Curve (AUC-ROC), Matthews Correlation Coefficient (MCC), and Cross Entropy Loss (CEL). MinMax + PCA + LSTM model has the best accuracy, AUC-ROC as well as MCC with the least loss, which shows that this model is robust and reliable. The model is verified on the publicly available GDM dataset in Kaggle proving its capabilities to be used in clinical decision support in GDM diagnosis.