Augmenting small biomedical datasets using generative AI methods based on self-organizing neural networks.

Area of Science:

• Biomedical research
• Computational biology
• Data science

Background:

• Small sample sizes are a significant limitation in biomedical research, hindering reproducibility and clinical translation.
• Challenges include limited resources, rare diseases, ethical constraints, and high diagnostic costs.
• Existing methods struggle to effectively augment small datasets without introducing bias.

Purpose of the Study:

• To propose a novel unsupervised generative algorithm to computationally increase sample sizes for small biomedical datasets.
• To address the challenges posed by limited data in fields like omics research.
• To improve the reliability and robustness of scientific findings from small or rare case studies.

Main Methods:

• A generative algorithm based on emergent self-organizing maps (ESOMs) was developed.
• The algorithm uses neural networks to identify data structure and generates new data points based on neighborhood probabilities.
• It adapts to high-dimensional data and generates synthetic samples that preserve original data characteristics.

Main Results:

• Experiments on artificial and biomedical (omics) datasets demonstrated that generated data preserve original structure without artifacts.
• Machine learning models (Random Forests, SVMs) could not differentiate between original and generated data.
• Statistical analysis confirmed no significant difference between variables of original and generated datasets.
• The method successfully augmented small datasets, including transcriptomics and lipidomics data.

Conclusions:

• The novel ESOM-based generative algorithm offers a promising solution for augmenting sample sizes in small or rare biomedical datasets.
• This approach can overcome limitations associated with small sample sizes, enhancing the reliability of research findings.
• The R library 'Umatrix' provides practical implementation of this method.