Addressing imbalance in health data: Synthetic minority oversampling using deep learning

Area of Science:

• Machine Learning
• Artificial Intelligence
• Healthcare Informatics

Background:

• Class imbalance in healthcare data leads to biased machine learning models, compromising patient safety and healthcare delivery.
• Traditional oversampling methods like SMOTE have limitations in handling complex data, heterogeneous types, and multi-class scenarios.

Purpose of the Study:

• To propose a novel deep learning approach for addressing class imbalance in healthcare datasets.
• To enhance the performance and fairness of machine learning models in clinical applications.

Main Methods:

• An Auxiliary-guided Conditional Variational Autoencoder (ACVAE) with contrastive learning was developed for synthetic data generation.
• An ensemble technique combining ACVAE for oversampling positive cases and Edited Centroid-Displacement Nearest Neighbor (ECDNN) for majority class undersampling was employed.

Main Results:

• Experiments on 12 health datasets demonstrated the effectiveness of the proposed ACVAE-ECDNN ensemble method.
• The approach showed significant improvements in model performance across various metrics compared to traditional oversampling techniques.

Conclusions:

• Deep learning-based synthetic oversampling offers a powerful solution for class imbalance in healthcare data.
• The proposed method enhances dataset balance and informativeness, leading to more reliable machine learning models for healthcare.