Scalable cloud–AI architecture for synthetic healthcare data generation and anomaly modeling: a simulation-based study

Synthetic healthcare data has emerged as a promising solution for enabling data-driven public health research while preserving patient privacy. However, existing generative frameworks often struggle with training stability, limited anomaly representation, and unclear scalability claims. This study presents a hybrid Variational Autoencoder–Generative Adversarial Network (VAE–GAN) integrated with a Reinforcement Learning (RL)–based anomaly injection mechanism for synthetic healthcare data generation. All model development, training, and evaluation were conducted using a MATLAB-based simulation environment. The cloud-native microservice architecture discussed in this work is presented as a conceptual deployment framework, rather than an empirically validated cloud implementation. To address limitations of purely synthetic validation, the proposed model was additionally evaluated against a subset of real-world anonymized clinical data from the MIMIC-III v1.4 database, focusing on distributional similarity, correlation preservation, and anomaly realism. Clinically meaningful anomalies were defined using established physiological thresholds, and the RL reward function was reformulated with a formal justification and supported by an ablation study examining exploration–exploitation behavior. Comparative evaluation was expanded to include diffusion-based, flow-based, and tabular GAN models. The results demonstrate that the proposed hybrid framework produces statistically consistent synthetic records with improved anomaly coverage, while maintaining privacy separation. This simulation-based study provides a rigorous methodological foundation for future real-world cloud deployment and public health applications.