A Novel Active Learning Technique for Fetal Health Classification Based on XGBoost Classifier

Abstract

Ensuring safe pregnancy and reducing maternal and infant mortality rates require early prediction of fetal health. The application of machine learning algorithms in monitoring fetal health helps to improve the chances of timely intervention and better outcomes in case of any possible health issues in fetuses. Existing studies offered to aid this issue typically by training models using a significant portion of the dataset, ranging mostly around above 70%. The only existing active learning method in this field employs around 41% training samples to achieve 98% accuracy. This work presents a novel active learning technique to identify the most informative data samples for training a model leading to high accuracy with limited number of training samples. It employs a novel query function built upon uncertainty and diversity criteria which are derived based on properties of XGBoost classifier and distance from each other. For deriving uncertainty criterion the soft probabilities obtained for the unlabelled samples are used while the distance among the uncertain samples in feature space is utilized for deriving diversity criterion. The proposed approach shows superior performance in comparison to all the state-of-the-art methods. Through analysis and experimentation, the proposed solution achieves an accuracy above 99% by utilizing less than 20% of the dataset for training. This demonstrates its efficacy and potential in fetal health monitoring. The code and dataset is available at the GitHub repository https://github.com/niyg7/fetal-health-dataset.

Topics & Concepts