Title:
Potential Applications of Transfer Learning in Limited Biomedical Data
Presenter:
Youngjun Park (FAIrPaCT Member)
Collaborators:
Anne-Christin Hauschild (FAIrPaCT Member)
Dominik Heider
Event:
GMDS 2022 – Annual Meeting of the German Association for Medical Informatics, Biometry and Epidemiology
Overview:
At GMDS 2022, Youngjun Park presented innovative research on how transfer learning can address key challenges in biomedical data analysis, especially in domains where data is limited—such as rare diseases or single-cell sequencing.
Key Insights:
- Transfer learning models, pre-trained on large datasets (e.g., GTEx and TCGA), can significantly improve performance on smaller, specialized datasets.
- The approach improved tissue classification across cancer types (TCGA) and enhanced cell-type classification in pancreas single-cell datasets.
- Fine-tuning with only 5% of target data reached accuracy levels comparable to state-of-the-art methods using 80% of data, reducing both resource demand and training time.
Why It Matters:
This work demonstrates that leveraging large public datasets can overcome batch effects, technological heterogeneity, and limited sample sizes often found in biomedical research. It highlights how transfer learning offers a promising path to robust, efficient analysis in healthcare AI, aligning closely with FAIrPaCT’s mission of advancing ethical, privacy-preserving machine learning.