Structural preprocessing and feature alignment for optimizing CTGAN-based synthetic generation from South African Global Health Observatory data

Abstract

High-fidelity synthetic data generation is a critical prerequisite for addressing data scarcity and privacy constraints in global health. While Conditional Tabular Generative Adversarial Networks (CTGANs) are established in synthesising patient-level electronic health records (EHRs), their performance on aggregate, high-variance transnational health indicators remains underexplored. This study introduces the Structural Preprocessing and Feature Alignment (SPFA) Framework, a novel, preprocessing-centric pipeline designed to optimise CTGAN performance for aggregate indicators. By evaluating the efficacy of the SPFA Framework in synthesising WHO Global Health Observatory data for South Africa, the research shifts the generative modelling focus from clinical settings toward the macro-level health surveillance domain.

The SPFA Framework implements a five-stage hierarchical pipeline incorporating automated type-safe null handling, per-indicator IQR-based outlier capping, dual-criterion cardinality reduction, and Log1p normalisation to mitigate heteroskedasticity. This approach was evaluated through a multi-criterion fidelity assessment, integrating SDV structural quality scores, statistical descriptor proximity, visual density histograms, and Pearson correlation preservation. The SPFA-optimised CTGAN achieved acceptable-to-good statistical fidelity across all dimensions, demonstrating strong marginal distributional alignment (SDV Column Shape ≥ 0.85) and accurate preservation of inter-feature relationships (Pearson correlation difference < 0.02). Notably, the mean and standard deviation of synthetic values differed from real-world data by only -5.81% and +1.27%, respectively. While minor divergence was observed in the upper quartile distributions, the pipeline successfully addressed the high cardinality and skewness inherent in South African health indicators. These results validate the SPFA Framework as a robust, reproducible methodology for public health. By demonstrating that structural alignment is a primary determinant of synthesis quality, comparable in influence to hyperparameter tuning, this study provides a scalable pathway for enhancing data utility in global health surveillance and evidence-based policy modelling.

Received 29 July 2026

Accepted 24 September 2026

Published 04 October 2026