A CVAE FRAMEWORK FOR THE GENERATION AND EVALUATION OF BRAIN TABULAR DATA

Medical datasets commonly suffer from a critical limitation: a significant imbalance in age representation, resulting in models biased towards the most represented age groups. This is particularly relevant in the context of brain aging research, where the Brain Age Gap (BAG) represents a fundamental biomarker of neurological health, measuring subject-specific deviations from population-based normative aging trajectories. Furthermore, another relevant application is normative learning, where healthy population trends are described as a function of covariates such as age and sex; it is therefore fundamental to have a sufficient number of subjects per group to adequately characterize not only the mean and standard deviation, but also higher-order moments of the underlying feature distributions, particularly in less-represented age groups. In this work, a novel framework is presented, designed to generate synthetic brain tabular data. The aim is to augment the existing datasets with synthetic samples for the most underrepresented age groups in order to obtain more stable and generalizable models across all age ranges. A Conditional Variational Autoencoder (cVAE), conditioned on Age and Sex, is employed as the generative architecture. In the presented framework, known VAE-based model limitations are addressed, such as the tendency to reproduce feature distributions as Gaussian, which fails to preserve higher-order statistical moments, a characteristic element of brain features. Moreover, the VAE-based model suffers from prior hole, a condition in which due to the mismatch between the prior distribution and aggregate approximate posterior, there exist areas in the latent space with high density under the prior, that do not correspond to any encoded sample. To this end, different types of covariate parameterizations are tested, serving as baseline architectures. Building on these, an ablation study is conducted by retraining each configuration with the addition of new loss terms: a moment loss, accounts for higher statistical moments, namely skewness and kurtosis, and correlation loss, which guarantees that most informative correlation patterns related to features most correlated with age are retained. To address the prior hole problem, a posterior sampling pipeline is developed, conditioning the sampling procedure on the subject's age range and sex according to predefined stratification criteria. Several experiments are conducted to validate the generated samples against the standard prerequisites for synthetic data quality: fidelity, diversity, and utility through a downstream task. The synthetic data generated using the modified versions of the training loss prove to be valid and improve the specific objectives they target, namely better preserving the skewness, kurtosis, and correlation patterns of the original features. These configurations maintain downstream task performance comparable to the baseline models, as well as a Random Forest discriminator AUC indicating that synthetic and real samples remain difficult to distinguish, consistent with results reported in the literature for similar medical applications. In conclusion, the developed cVAE framework proved effective for generating synthetic data that satisfy the required criteria, and the various loss terms can be selectively applied depending on the intended downstream task (e.g., normative learning or data augmentation).