Evaluating Synthetic Data Augmentation to Correct for Data Imbalance in Realistic Clinical Prediction Settings

Area of Science:

• Medical Informatics
• Machine Learning
• Data Science

Background:

• Clinical decision-making relies on predictive modeling, often challenged by imbalanced datasets.
• Data augmentation is crucial for improving model performance on underrepresented classes.

Purpose of the Study:

• To evaluate the effectiveness of synthetic data generation for enhancing predictive models on small, imbalanced clinical datasets.
• To compare advanced synthetic data techniques against traditional methods for class imbalance correction.

Main Methods:

• Investigated Generative Adversarial Networks (GANs), Normalizing Flows, and Variational Autoencoders (VAEs).
• Compared these methods to standard baselines for class underrepresentation.
• Utilized four realistic clinical datasets for evaluation.

Main Results:

• Some synthetic data methods showed marginal improvements in F1 scores.
• No statistically significant evidence was found that synthetic data generation outperformed standard baselines.
• Results were consistent even after multiple repetitions.

Conclusions:

• The efficacy of synthetic data for augmenting imbalanced clinical data requires careful evaluation.
• Complex synthetic data generation methods may not consistently outperform simpler, established techniques.
• Emphasizes the need to benchmark novel methods against standard baselines in predictive modeling research.