JMIR Publications

ABSTRACT

Background:

Synthetic data generation (SDG) based on generative adversarial networks (GANs) has garnered significant attention in healthcare in the context of various tasks. However, minimal research has been conducted on SDG that preserves logical relationships and appropriates synthetic tabular data (STD) for learning in healthcare. Several researchers have studied SDG based on filtering methods—however, record selection using such methods depends only on predefined condition columns, which may induce the exclusion of meaningful information.

Objective:

The purpose of this study is to propose a divide-and-conquer (DC) approach and use it to generate STD, which preserves logical relationships for model learning based on GAN algorithms.

Methods:

The DC-based SDG strategy comprises four primary components. First, we define the division criteria for training. The first criterion is “class-specific“, i.e., it depends on the class between survival and death groups. The second criterion uses the “Cramer’s V” correlation measure, which identifies the highest correlation between columns in the original data (OD). Subsequently, the entire dataset is divided into several subsets following the aforementioned definition. Then, CTGAN and CopulaGAN are trained on the two divided data subsets to generate synthetic data. Finally, the generated synthetic data are combined into a single entity in the conquer step. For validation, the prediction performances of decision tree (DT), random forest (RF), extreme gradient-boosting (XGB), and light gradient-boosting machine (LGBM) are compared with the proposed approach and the conditional sampling (CS) approach of CTGAN and CopulaGAN. Also, prediction performances are compared on balanced synthetic and imbalanced synthetic datasets.

Results:

The experimental results reveal that the proposed model exhibits more accurate prediction performance with respect to the OD than SDG generated using existing methods. DC-based synthetic data is higher quality than synthetic data produced via CS as per the classification methods; DT: CTGAN (DC: 74.5 ± 1.2 vs CS: 60.0 ± 1.3), CopulaGAN (DC: 74.9 ± 0.8 vs CS: 70.5 ± 0.8), and OD (66.1 ± 1.3); RF: CTGAN (DC: 85.6 ± 0.3 vs CS: 79.0 ± 1.2), CopulaGAN (DC: 83.9 ± 0.4 vs CS: 78.2 ± 1.7), and OD(84.8 ± 0.2); XGB: CTGAN (DC: 85.2 ± 0.8 vs CS: 74.7 ± 1.6), CopulaGAN (DC: 83.6 ± 0.7 vs CS: 76.4 ± 0.9), and OD (83.1 ± 0.4); LGBM: CTGAN (DC: 85.2 ± 0.6 vs CS: 77.8 ± 1.5), CopulaGAN (DC: 83.7 ± 0.5 vs CS: 77.6 ± 1.9, and OD (84.0 ± 0.0). Moreover, models with balanced STDs outperform those withoutperform those with imbalanced STDs.

Conclusions:

Besides being the first attempt to generate and validate STDs based on a DC approach while preserving logical relationships, this study demonstrates that the proposed method exhibits improved performance. The necessity for balanced synthetic data generation is also demonstrated.