Enhancing Predictive Healthcare Models through Counterfactual Data Augmentation: A Novel Approach Leveraging...

The scarcity of high-quality labeled data hinders the accuracy and reliability of healthcare prediction models, whereas data generation methods may enhance predictive models by augmenting training datasets. However, traditional data generation methods rely on imitating existing data patterns, limiting their ability to adapt to unforeseen circumstances. In this study, we propose a data generation method that does not rely on explicit causal structures and exploits cyclic structures to envision counterfactual states of medical data, thereby providing a diverse and extended dataset for predictive modeling. We introduce the cycle counterfactual residual generative adversarial network (CCR-GAN) and evaluate it on two datasets: a publicly available stroke prediction database with 4,886 patients and a hemodialysis (HD) dataset with 1,229 patients. The CCR-GAN integrates residual structures and a cyclic architecture, involving two key processes: generating counterfactuals for real instances across various clinical outcomes and reconstructing real instances from counterfactual scenarios. We compare the CCR-GAN with the state-of-the-art GAN models to assess its downstream prediction task improvements and the privacy of synthetic data. Additionally, we performed ablation studies to evaluate the utility, privacy, realism, and actionability of the counterfactuals. The data generated by the CCR-GAN model demonstrates significant predictive modeling accuracy improvements across the stroke dataset and the HD dataset, and shows competitive privacy protection performance compared to the baseline methods, achieving the lowest scores in attribute inference and membership inference tests. The results show that counterfactual data augmentation significantly enhances the accuracy of prediction models while preserving privacy and maintaining a certain degree of realism.