Large language models are now moonlighting as clinical consultants for AI systems that track rare diseases.
A new paper describes feeding LLMs text descriptions of patient visits offline and asking them to render a clinical judgment, such as the likely disease subtype. Those judgments train a smaller surrogate model, which then nudges the loss function of a variational autoencoder so that its compressed, low-dimensional summaries of patient data do not drift across clinical boundaries during reconstruction. Applied to longitudinal motor-function records from children with spinal muscular atrophy, the approach cut disagreement between original and reconstructed disease-type labels from about 11 percent to 7 percent. It also beat both unsupervised representations and a plain data-level baseline at predicting motor-function milestones.
Rare-disease datasets are small by definition, and the clinicians who could label them by hand rarely have spare time. This method effectively launders informal medical judgment into a model's training signal without a doctor annotating every record, trading the usual bottleneck of expert hours for compute time instead.
That trade is only as good as the LLM doing the judging. The "synthetic expert" here is a language model inferring diagnoses from text descriptions, not a clinician examining a patient, and any blind spots or biases in its medical reasoning get baked directly into the representation learning. A useful workaround for data-starved fields, but one worth watching closely before it moves past research prototypes.