CEDAR, a self-checking chain of LLM agents, is ahead of machine learning on paediatric type 1 and 2 diabetes cases, with only 4 erroneous quotes out of over 70,000.
Agents based on large language models are often proposed for clinical prediction, but paediatrics pose a particular problem: small cohorts and rare events. A team of 11 researchers tested CEDAR (Clinical Evidence-Driven Agent for Risk-prediction) for 2,489 youth with type 2 diabetes and 4,018 with type 1 diabetes. The system reads structured data, social factors extracted from medical notes and, for type 1, continuous glucose sensor readings, to predict complications at 24 months from 18 months of follow-up.
More efficient, and verifiable
Of seven criteria for type 2 and four for type 1, CEDAR achieves an area under the mean curve of 0.815 and 0.756, compared with 0.60 to 0.70 for specifically trained machine learning models. The gap is the clearest on rarely detected complications. Important point for clinical adoption: the officer justifies each prediction by quoting elements of the record, and only 4 of the 70,246 values cited did not match the record. Two paediatric endocrinologists reading only these justifications reached the same conclusion as the agent in approximately nine out of ten cases.
Read the full article on the original source
Go to the source →