Knowledge Graph Modulated Deep Learning for Limited-Sample Clinical Data Analysis
Researchers have developed a new deep learning framework called Graph-in-Graph (GiG) designed to improve clinical data analysis, particularly in situations with limited patient samples. GiG integrates biological knowledge graphs directly into the patient representation learning process, preserving crucial gene-gene interactions and pathway topology. Across five clinical tasks and nearly 9,700 patients, GiG demonstrated superior performance compared to existing methods, showing significant gains in sample efficiency and accuracy, such as a 49 percentage point improvement in macro-F1 for prostate cancer diagnosis. AI
IMPACT Enhances sample efficiency and accuracy in clinical AI, particularly for limited-data scenarios.