MOSAIC scored 95.6% in clinical communication coding
The multi-agent system was tested on 120 transcripts and was strongest in rheumatology.
MOSAIC is an AI system that codes clinical conversations across multiple communication frameworks, and a paper in npj Digital Medicine reported that it reached a 95.6% weighted F1 score.
The system was built for the slow, messy job of turning patient-provider talks into structured labels. That work usually takes trained human coders a lot of time.
MOSAIC reached an overall weighted F1 score of 95.6%.
MOSAIC combines three pieces: dynamically few-shot prompting, retrieval-augmented generation, and a LangGraph-based multi-agent workflow. In plain terms, it uses examples on the fly, pulls in relevant reference material, and splits the task across planning, annotation, and verification steps.
The team developed it on 54 annotated transcripts.
They then tested it on 120 transcripts from rheumatology and obstetrics/gynecology, comparing the output with gold-standard labels made by human coders.
The overall weighted F1 score came out at 95.6%. That is a measure of how closely the system matched the reference annotations.
Performance was highest in rheumatology, where the F1 score reached 95.9%.
Two categories stood out as especially strong: bias at 96.4% and patient behaviors at 96.1%.
Imagine having to sort hundreds of clinic conversations by hand. MOSAIC is trying to be the fast sorter that still double-checks its own homework.
The paper shows strong results on the tested transcripts, but it does not show performance beyond rheumatology and OB/GYN, or beyond the 120-transcript evaluation set.
Source: npj Digital Medicine