Dimensionality-modulated generative AI for safe biomedical dataset augmentation

Generative AI can expand small biomedical datasets but may amplify noise and distort statistical relationships. We developed genESOM, a framework integrating an error control system into a generative AI method based on emergent self-organizing maps. By separating structure learning from data synthesis, genESOM enables dimensionality modulation and injection of engineered diagnostic features, i.e., permuted versions of real variables, as negative controls that track feature importance stability. A data-driven stopping criterion halts augmentation when error inflation begins, limiting overfitting. Validation across two artificial and six biomedical datasets, including preclinical and clinical domains, showed that moderate augmentation (one synthetic sample per original) preserved variable ranking and yielded strong negative correlations (Kendall's tau: -0.53 to -0.85) between statistical significance and feature selection frequency. Excessive augmentation disrupted these relationships. Moderate preclinical augmentation safely doubled sample sizes without compromising analytical reliability and supported up to 50% reductions in laboratory animal use.