Hybrid GAN with KNN - SMOTE Approach for Class-Imbalance in Non-Invasive Fetal ECG Monitoring

Abstract

Imbalanced fetal electrocardiogram (fECG) datasets often hinder reliable fetal health assessment by biasing predictions toward majority classes. This article presents a two-stage augmentation framework that integrates three generative models: conditional GAN (cGAN), time-series GAN (TGAN), and Wasserstein GAN (WGAN) with a K-nearest neighbor (KNN) - based Adaptive Synthetic Minority Over-Sampling Technique (SMOTE) algorithm to generate physiologically realistic minority-class signals. Preprocessing steps included the removal of missing records and normalization to ensure data consistency. The balanced dataset was used to train a hybrid LSTM–CNN classifier designed to capture both long-term temporal dynamics and localized time–frequency features of fECG signals. The proposed method improved overall classification accuracy by 4–7% and minority-class F1-scores by up to 10% compared to baseline approaches. The framework achieved 97% accuracy and 98% F1-score by combining ensemble GAN-based augmentation with adaptive oversampling for robust and balanced biomedical time-series analysis.