EXPLORING ADVANCED MACHINE LEARNING TECHNIQUES FOR EFFICIENT PREDICTION OF PRETERM BIRTH

Abstract

This work's purpose is to contribute to the field of premature birth prediction with the help of Electrohysterogram (EHG) signals. Since preterm birth is critical public and health challenge, the project aims to find a reliable way to deal with the issue of class imbalance - one of the problems in prediction - by comparing various oversampling, undersampling, and ensemble techniques, with and without Cross-Validation (CV). Additionaly, exploration of different features and their effect on the model performance has been conducted. The results have shown that using stratified 5-fold CV in conjunction with resampling approaches might be a useful strategy for managing unbalanced datasets and producing more accurate performance measures. Resampling techniques can aid in "balancing" the dataset, but their effectiveness is best evaluated with CV, which guarantees that the model's performance accurately represents its generalization. Moreover, the model's feature type has a big impact on performance: categorical features showed superior results without CV, but when CV has been implemented, performance became more consistent across feature sets.