Exploring the Joint Effect of Resampling Techniques and Cross-Validation for Preterm Birth Prediction

Area of Science:

• Medical Informatics
• Machine Learning
• Biomedical Engineering

Background:

• Imbalanced datasets are prevalent in real-world classification tasks, notably in predicting preterm birth.
• Standard machine learning algorithms often exhibit bias on imbalanced data, impacting predictive accuracy and fairness.
• Preterm birth prediction is challenged by the rarity of preterm cases, necessitating robust data handling methods.

Purpose of the Study:

• To systematically compare various undersampling, oversampling, and ensemble techniques for addressing class imbalance in preterm birth prediction.
• To evaluate the impact of Cross-Validation (CV) on the performance of these resampling methods.
• To identify the most effective resampling strategies for improving the accuracy and reliability of preterm birth prediction models.

Main Methods:

• Utilized the Term-Preterm Electrohysterogram (EHG) Database (TPEHG DB) for analysis.
• Implemented and compared multiple undersampling, oversampling (e.g., Synthetic Minority Oversampling Technique - SMOTE), and ensemble techniques.
• Assessed the performance of these methods both with and without Cross-Validation (CV).

Main Results:

• Ensemble techniques demonstrated the most consistent performance across different resampling strategies.
• Specific methods like SMOTE, Cluster-Based Undersampling (Cluster Centroids), and Ranked Minority Oversampling in boosting (RAMOBoost) yielded superior overall results.
• The study addressed previous research limitations concerning data handling and overly optimistic outcomes.

Conclusions:

• The findings suggest that ensemble methods are highly effective for managing class imbalance in preterm birth prediction.
• The identified techniques (SMOTE, Cluster Centroids, RAMOBoost) offer promising solutions for developing more accurate prognostic tools.
• This research represents a foundational step towards creating reliable resampling tools with significant clinical relevance for preterm delivery prognosis.