Self-organizing neural network-based generative AI with embedded error inflation control enhances effective knowledge...

Area of Science:

• Biomedical research
• Artificial intelligence
• Data science

Background:

• Small sample sizes in preclinical studies limit reliable knowledge extraction and translational progress.
• Current methods struggle to augment limited biomedical data effectively while controlling statistical errors.

Purpose of the Study:

• To introduce genESOM, a novel generative AI method for augmenting small biomedical datasets.
• To demonstrate genESOM's ability to control alpha-error inflation and enable safe, interpretable data augmentation.
• To evaluate genESOM's performance against other generative models using a preclinical multiple sclerosis dataset.

Main Methods:

• Developed genESOM, a generative AI method based on emergent self-organizing maps, separating structure learning from data synthesis.
• Integrated error propagation mitigation via dimensionality modulation for controlled data augmentation.
• Applied genESOM to a reduced dataset (18 mice) from a preclinical EAE multiple sclerosis study (originally 26 mice, 62 lipid mediators).

Main Results:

• Reducing sample size from 26 to 18 mice abolished detectable group differences in statistical and machine learning analyses.
• Augmenting the reduced dataset with genESOM-generated cases restored treatment-specific segregation and identified key lipid mediators.
• genESOM achieved high fidelity without false positives, outperforming Gaussian mixture and conditional GAN models under similar constraints.

Conclusions:

• genESOM offers a robust, error-controlled framework for enhancing knowledge extraction from limited preclinical data.
• AI-driven synthetic data augmentation can support exploratory analyses and potentially reduce the need for animal experimentation.
• genESOM demonstrates a viable strategy for overcoming data limitations in preclinical research.