Authors release RetiJev models for selecting answers in eye diagnosis and clinical decision-making
The RetiJev family of models, with 0.8–9 billion parameters, directly scores answers prepared in advance. According to the authors, RetiJev-9B outperforms two comparison systems by 5.4–6.5 points in clinical decision-making and processes a case in 158 ms on a single GPU.
The authors released the RetiJev family of models to support decision-making in ophthalmology. The models have 0.8–9 billion parameters and score candidate answers directly without autoregressive generation. They focus on diagnosis, clinical decisions and classification according to clinical guidelines.
RetiJev-Data contains 78 158 clinical decision-making and diagnostic questions derived from articles in PubMed Central and the MeSH vocabulary covering 531 eye diseases. A further 50 074 questions concern 11 classification systems. A separate evaluation set contains 24 662 questions.
According to the authors, the RetiJev models outperform both comparison systems, an LLM with 27 billion parameters and the commercial model Jev 1.13, by 10.4–23.0 points in classification according to guidelines. In their measurements, RetiJev-9B leads by 5.4–6.5 points in clinical decision-making, while achieving comparable results in diagnosis. The authors report processing a case in 158 ms on a single GPU, which is 3.5 times faster than generation using an LLM and 470 times faster than a reasoning-based approach.
Why it matters
For developers building tools that select from predefined clinical options, the released model family offers a concrete way to reduce response time. The reported accuracy and speed come from evaluations by the authors; on their own, they do not establish safety or benefits in patient care.
Relevant practical impact
What this means
For individuals
A developer or researcher in ophthalmology can try the released models during prototyping to score candidate answers directly.
Check the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.