Synthetic data generation from small real-world datasets for biological age estimation

Abstract

This article addresses the relevance of overcoming the shortage of high-quality clinical data required for training reliable machine learning models for biological age estimation. The collection of such data is expensive, time-consuming, and restricted by privacy regulations, while small sample sizes lead to overfitting. The aim of this research is to compare modern synthetic data generation methods using a limited set of skeletal system biomarkers to identify the optimal approach under small-sample conditions. The object of the study is the methodology of generating synthetic data for small tabular biomedical records applied to estimating biological age. The subject of the study encompasses statistical methods and generative machine learning models used to synthesize realistic clinical information. The objectives include selecting relevant biomarkers, implementing various data generation frameworks, assessing the quality of synthetic data, and evaluating privacy risks. The study evaluates the bootstrap resampling method, the synthetic minority oversampling technique, conditional tabular generative adversarial networks, tabular variational autoencoders, and the Gaussian copula approach. Validation relies on nonparametric statistical tests for distributional similarity, geometric proximity measurements, nearest neighbor privacy evaluations, and cross-validation testing. The results demonstrate that basic resampling fails to protect data privacy, while linear oversampling severely distorts chronological age distributions and lacks sample uniqueness. Deep generative adversarial networks exhibit mode collapse under small-sample conditions, shown by significant error discrepancies between training and testing scenarios. In contrast, the statistical copula approach provides the highest fidelity in maintaining complex multivariate correlations. Additionally, probabilistic autoencoders prove exceptionally capable of reproducing marginal distributions. Both of these successful methods effectively maintain privacy. The scientific novelty lies in the development of a comprehensive comparative framework for evaluating generative models under strict small-data constraints. The practical value of the study consists in providing clear recommendations for researchers: the Gaussian copula approach should be prioritized when preserving exact physiological relationships between features is critical, whereas tabular variational autoencoders are optimal when accurate population-level distributions are strictly required.