MedKGTab Expands Medical Data Features Using Knowledge Graphs.

Researchers have developed MedKGTab, a novel framework designed to overcome data scarcity in medical research by expanding features in tabular medical data. This system infers missing biomedical features from existing data by combining statistical dependencies with established medical correlations found in the SPOKE biomedical knowledge graph. MedKGTab employs a row-column dual-attention mechanism, allowing it to operate directly on raw structured tabular data and preserve exact numerical distributions without the loss associated with tokenization. A key innovation of MedKGTab is its ability to integrate data-driven statistical priors with external biomedical knowledge, ensuring that the generated data is empirically grounded. Experimental results demonstrate that MedKGTab achieves superior data fidelity and realistic representation in cross-domain feature expansion. It consistently outperforms both leading medical large models, such as Baichuan M3-plus, and specialized tabular data generation models across various scenarios, including inferring missing features within a dataset and generalizing across different medical cohorts.