Abstract
Imbalanced datasets are common in real-world classification problems, such as predicting premature birth. Many machine learning algorithms, designed for balanced data, may produce biased predictions when applied to imbalanced datasets. In preterm birth prediction, where preterm cases are much less frequent than full-term, this imbalance may significantly impact model performance. Resampling techniques like oversampling the minority class or undersampling the majority class are key to improving accuracy and fairness. Previous studies have struggled with data handling, often leading to overly optimistic results, which this paper aims to address. The main objective is to compare various undersampling, oversampling, and ensemble techniques, with and without Cross-Validation (CV). For this purpose, the Term-Preterm Electrohysterogram (EHG) Database (TPEHG DB) has been used. Results show that ensemble techniques perform the most consistently. It has been found that the Synthetic Minority Oversampling Technique (SMOTE), Cluster-Based Undersampling (Cluster Centroids), and Ranked Minority Oversampling in boosting (RAMOBoost) yield the best overall performance.Clinical relevance - The present work may be considered as the first step in developing robust resampling tool for dealing with class imbalance in the prediction of preterm birth. The preliminary results indicate that machine learning and resampling techniques can aid in the research of preterm delivery prognosis, which may have significant clinical relevance for both healthcare providers and patients.