In the high-stakes world of ophthalmology, clinical decision support (CDS) needs to be three things: fast, accurate, and private. While Large Language Models (LLMs) have shown impressive medical knowledge, their "generative" nature—predicting one token at a time—makes them frustratingly slow and computationally expensive for point-of-care use.
RetiJev addresses this by shifting the paradigm from generation to scoring. Instead of writing a long-winded explanation, RetiJev models directly evaluate a set of predefined clinical actions, providing near-instantaneous, source-grounded decisions that can run on a single GPU or even a standard CPU.
TL;DR
RetiJev is a family of compact (0.8B–9B) decision models specialized for ophthalmology that outperform 27B LLMs while being up to 470x faster. By scoring candidate answers directly rather than generating text, they provide highly accurate, guideline-adherent clinical support with low enough latency for real-time use during patient encounters.
What Problem Does This Paper Address?
Clinical decision-making in ophthalmology is often time-critical (e.g., central retinal artery occlusion has a treatment window measured in hours). However, specialists are scarce, and many initial calls are made by non-specialists.
Current LLMs struggle in this environment for several reasons:
- Latency: Autoregressive decoding is slow, especially when using "Chain-of-Thought" reasoning.
- Computational Cost: Large models require massive GPU memory, making local, private deployment in a clinic difficult.
- Guideline Inconsistency: Generative models often hallucinate or fail to strictly follow established clinical grading systems (like diabetic retinopathy scales).
The authors argue that for most clinical tasks—like determining urgency, selecting the next investigation, or grading a disease—we don't need a model to talk; we need it to choose correctly from a set of valid medical options.
Methods
RetiJev moves away from the standard "chat" interface to a Single-Pass Decision Model architecture.
- Architecture: The models use a Qwen3.5 hybrid backbone (combining standard attention with Gated DeltaNet layers). Instead of a language modeling head, they use a shared trainable pointer head. This head projects the patient state and candidate answers into a shared space to calculate probabilities for each option in a single forward pass.
- Shared-Context Inference: In a clinical setting, a doctor might ask multiple questions about one patient (e.g., "What is the grade?" and "How urgent is this?"). RetiJev encodes the patient's clinical note once and reuses that representation to score all questions in parallel, drastically reducing redundant computation.
- Training Objectives: The models are trained using a mix of hard label supervision, teacher distillation (learning from a larger 27B model's probability distributions), and paraphrase consistency to ensure the model isn't confused by different ways of writing the same clinical finding.
Datasets
The researchers introduced RetiJev-Data, a massive specialized corpus:
- Clinical-Decision Resource: 78,158 questions derived from PubMed Central articles, covering 531 eye diseases.
- Guideline-Grading Resource: 50,074 questions based on 11 published grading systems (e.g., ICDR for diabetic retinopathy).
- RetiJev-Bench: A 24,662-question evaluation benchmark designed to test models on both "seen" topics and entirely "unseen" clinical rules.
Key Insights and Ideas
- Specialization Beats Scale: A 0.8B parameter model, when properly specialized for decision-scoring, can match or exceed the performance of a 27B general-purpose LLM on domain-specific tasks.
- The "Speed is Everything" Commandment: By eliminating the decoding step, the authors achieved a "reasoning-level" accuracy at "system-one" speeds.
- Robustness through Paraphrasing: One of the biggest hurdles for small models is sensitivity to wording. The authors found that explicitly training for paraphrase consistency was the "secret sauce" that allowed their compact models to remain accurate across diverse clinical note styles.
Key Results
RetiJev sets a new bar for efficiency and accuracy in ophthalmic AI:
- Accuracy: RetiJev-9B outperformed a 27B LLM and the commercial Jev 1.13 model on guideline grading by 10.4–23.0 percentage points.
- Speed: RetiJev-9B answers a case in 158 ms, which is 3.5x faster than standard LLM generation and 470x faster than reasoning-based generation (which took over a minute per case).
- Efficiency: The 0.8B model requires only 3.3 GB of GPU memory and can run on 8 CPU cores in just 0.35s, making it viable for deployment on basic hospital hardware without cloud dependencies.
Limitations and Future Work
While the results are impressive, the authors note several caveats:
- Synthetic Data: The patient cases used for training and testing are synthetically generated (though grounded in real-world literature and verified by other models).
- Clinical Validation: The model is a research prototype and has not yet undergone prospective clinical trials or clinician-adjudicated evaluation on real-world patient records.
- Disease Recognition: While excellent at decisions and grading, smaller models (0.8B) still lag behind massive LLMs in recognizing very rare diseases, suggesting that model scale still matters for broad diagnostic knowledge.
Future work will likely focus on integrating multimodal data (like retinal images) and conducting real-world pilot studies in clinical environments.
Summary updated