Graph neural networks for networked analysis of gestational diabetes risk factors: a multi method framework

Description

Lella and team modeled diabetes that arises in pregnancy as a network problem, linking patients who resemble each other instead of scoring each woman in isolation. Gestational diabetes risk depends on tangled genetic, biochemical, and demographic factors that plain forecast models flatten. Their 2026 design builds a graph where patients are nodes, edges mark shared traits, and graph neural networks pass information along those edges, while a language model reads the unstructured clinical notes.

The combined model reached about 88 percent classification accuracy on the clinical dataset evaluated, beating the baseline learning methods compared. Three graph designs, covering neighborhood sampling, relation-aware messages, and time-aware attention, each contributed to the result. Privacy-preserving federated training and edge-dropping regularization guarded against overfitting and data exposure. Lella argues network context, the company a patient keeps in data, sharpens risk forecast.

Catalog record

Source: https://europepmc.org/articles/PMC13490363