Synthetic Data Augmentation

TVAE was considered the strongest augmentation method because it better preserved clinical feature distributions and generated realistic synthetic patient profiles. SMOTE created additional minority-class readmission cases by interpolating between observed samples. TVAE used a variational autoencoder to learn relationships in EHR data and generate realistic synthetic patient records. CTGAN learned patterns from real tabular clinical data and generated synthetic records while preserving mixed variable types and distributions. Synthetic data generation occurred only inside training datasets after partitioning to avoid validation leakage. The study tested SMOTE, CTGAN, and TVAE to address outcome imbalance.