JMIR Publications

ABSTRACT

Background:

Synthetic data generation (SDG) using generative adversarial networks (GANs) is used in healthcare, but research on preserving data with logical relation with synthetic tabular data (STD) has remained challenging. Previous studies have explored filtering methods for SDG, which can lead to the loss of important information.

Objective:

This study proposes the Divide-Conquer(DC) method to generate STD based on the GAN algorithm, while preserving data with logical relation.

Methods:

The DC-based SDG strategy comprises three steps. In the first step, we use two different partitioning methods. "Class-specific" distinguishes between survival and death groups. In addition, Cramer's V identifies the highest correlation between columns in the original data. In the second step, the entire data set is divided into a number of subsets, which are then used as input for CTGAN and CopulaGAN to generate synthetic data. In the third step, the generated synthetic data is consolidated into a single entity as the conquer step. For validation, we compared DC-based SDG and Conditional Sampling(CS)-based SDG through the performances of machine learning models. CS is a method used in CTGAN and CopulaGAN. CS is achieved through the use of rejection sampling, where rows are sampled repeatedly until the desired condition is met. The performance is also compared on balanced and unbalanced synthetic datasets.

Results:

The proposed method is evaluated on data from The Korea Association for Lung Cancer Registry (KALC-R) and two benchmark datasets (breast and diabetes). We compared the synthetic data generated by CTGAN or CopulaGAN to see which resulted in the highest performance. As a result, the synthetic data of the three diseases generated by our proposed model outperformed the four classifiers. Using a synthetic dataset of CTGAN and CopulaGAN with CS or DC, we compared the best model performance based on the area under the curve; decision tree (KALC-R, 74.87±0.77 (DC) vs 63.87±2.02(CS); breast, 73.31±1.11 (DC) vs 67.96±2.15(CS); diabetes, 61.57±0.09 (DC) vs 60.08±0.17(CS)), random forest (KALC-R, 85.61±0.29 (DC) vs 79.01±1.20(CS); breast, 78.05±1.59 (DC) vs 73.48±4.73 (CS); diabetes, 59.98±0.24 (DC) vs 58.55±0.17(CS)), extreme gradient-boosting (KALC-R, 85.20±0.82 (DC) vs 76.42±0.93(CS); breast, 77.86±2.27 (DC) vs 68.32±2.37(CS); diabetes, 60.18±0.20 (DC) vs 58.98±0.29(CS)), and light gradient-boosting machine (KALC-R, 85.14±0.77 (DC) vs 77.62±1.85(CS); breast, 78.16±1.52 (DC) vs 70.02±2.17(CS); diabetes, 61.75±0.13 (DC) vs 61.12±0.23(CS)). We also generated unbalanced and balanced synthetic data for each of the three datasets and compared their performance using DT, RF, XGBoost, and LGBM models, and found that balanced synthetic data performed better.

Conclusions:

Besides being the first attempt to generate and validate STDs based on a DC approach and shows improved performance. The necessity for balanced synthetic data generation is also demonstrated.