Dimensionality-modulated generative AI for safe biomedical dataset augmentation.

Area of Science:

• Biomedical data science
• Artificial intelligence in healthcare
• Machine learning for biological data

Background:

• Generative AI offers potential for expanding small biomedical datasets.
• However, generative AI may introduce noise and distort statistical relationships.
• Existing methods lack robust control over data augmentation to prevent overfitting.

Purpose of the Study:

• To develop a novel framework, genESOM, for controlled generative AI-based biomedical data expansion.
• To integrate an error control system and diagnostic features for reliable data synthesis.
• To validate genESOM's performance across diverse artificial and biomedical datasets.

Main Methods:

• Developed genESOM, integrating error control into emergent self-organizing maps for data synthesis.
• Separated structure learning from data synthesis, enabling dimensionality modulation.
• Incorporated engineered diagnostic features (permuted variables) as negative controls.
• Utilized a data-driven stopping criterion to prevent overfitting during augmentation.

Main Results:

• Moderate data augmentation (1:1 ratio) preserved variable ranking and statistical relationships across datasets.
• Strong negative correlations (Kendall's tau: -0.53 to -0.85) observed between statistical significance and feature selection frequency.
• Excessive augmentation disrupted these crucial analytical relationships.
• Preclinical data augmentation safely doubled sample sizes without compromising reliability.

Conclusions:

• genESOM provides a reliable method for expanding small biomedical datasets using generative AI.
• Controlled augmentation preserves analytical integrity and can reduce the need for extensive laboratory animal use.
• This framework supports more efficient and ethical preclinical research.