LLMs as Synthetic Clinical Experts for Rare-Disease Modeling

A recent preprint on arXiv (2608.16507) suggests leveraging large language models (LLMs) as synthetic clinical specialists to enhance the analysis of longitudinal data related to rare diseases. This technique merges clinical insights with a variational-autoencoder framework, which captures low-dimensional latent representations of visit-specific observations. LLMs are utilized offline to assess textual patient observation descriptions, providing insights like potential clinical categories. A differentiable surrogate model is developed based on these insights, with an augmented loss function that promotes reconstructions aligning with the clinical-label distribution of the original data. This innovative method tackles the issue of scarce expert availability and the formalization of knowledge necessary for model fitting, particularly in rare disease scenarios where integrating clinical expertise is advantageous. The full paper can be accessed on arXiv.