Collecting prospective observational clinical data is a complex and resource-demanding process often involving direct patient contact, repeated follow-ups, and specialized data collection procedures. These challenges typically limit sample size and increase susceptibility to model overfitting. Synthetic data generation is a promising approach for augmenting limited clinical tabular data and improving generalization. In this study, we implement and benchmark different augmentation strategies to improve the identification of chronic patients’ decompensation states based on physiological and clinical observations. The dataset comprises 117 patients and 284 visits, with repeated measurements capturing both decompensated and compensated states in heart failure or chronic obstructive pulmonary disease. Four data augmentation techniques were evaluated: a proposed Transformer-based Tabular Variational Autoencoder with Maximum Mean Discrepancy regularization and input masking (TTVAE-MMD), a Wasserstein GAN with gradient penalty, and two conventional interpolative and probabilistic approaches. Synthetic data quality was assessed using pyMDMA auditing library via (a) fidelity, (b) diversity, (c) authenticity, and (d) privacy metrics, along with analyses of marginal distributions and inter-feature correlations. TTVAE-MMD outperformed all other methods, achieving better diversity, fidelity, and privacy. Utility was assessed through the classification performance of a model optimised for each augmented subset, evaluated on entirely unseen individuals. The model trained on TTVAE-MMD augmented data reached the best F1-scores, representing a \(\sim \) 14% increase against real data baseline model. Findings show that high-quality synthetic data effectively augmented this small clinical dataset, improving model performance and generalization, with strong practical deployment potential.