Addressing Class Imbalance in Fetal Health Classification: Rigorous Benchmarking of Multi-Class Resampling Methods on...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Signal Processing for Biomedical Applications

Background:

• Fetal health monitoring via Cardiotocography (CTG) is crucial for prenatal care outcomes.
• Manual CTG interpretation suffers from significant inter-examiner variability.
• Class imbalance in CTG datasets (<10% pathological cases) hinders accurate detection of critical conditions.

Purpose of the Study:

• To systematically benchmark five resampling strategies against seven classifier families for multi-class CTG classification.
• To evaluate performance using imbalance-aware metrics beyond overall accuracy.
• To establish reliable, model-agnostic baselines for automated fetal health screening.

Main Methods:

• Seven machine learning models (e.g., Random Forest, SVM, MLP) were evaluated.
• Five resampling methods (SMOTE, BSMOTE, ADASYN, NearMiss, SCUT) were compared against the original imbalanced dataset.
• Performance was assessed using Balanced Accuracy (BACC), Macro-F1, Macro-MCC, and Macro-Averaged ROC-AUC.

Main Results:

• Random Forest (RF) emerged as the most reliable model.
• RF with BSMOTE achieved the strongest class-balanced performance (Macro-MCC = 0.8533, Macro-F1 = 0.9073).
• Oversampling techniques, especially SMOTE and BSMOTE, significantly improved minority class discrimination and balanced metrics across models.

Conclusions:

• Strategic oversampling effectively enhances minority class detection in CTG analysis.
• These findings provide robust baselines for deploying machine learning in fetal health screening.
• Improved discrimination translates to clinically meaningful performance gains in cost-sensitive settings.