P.121 Bridging the evidence gap: RAG-enabled LLMs in neuroimaging decision support
Bibliographic record
Abstract
Background: Large language models (LLMs) offer potential for clinical decision support but may not fully adhere to current guidelines. Retrieval-augmented generation (RAG) may address this gap by dynamically incorporating external knowledge. This study evaluated LLM adherence with and without RAG to Canadian neuroimaging guidelines. Methods: A novel RAG framework was developed that integrated Canadian Association of Radiologists (CAR) Diagnostic Imaging Referral Guidelines with GPT-4o and o1 models. Clinical scenarios were curated to represent various central nervous system conditions, such as acute stroke, subarachnoid hemorrhage, and multiple sclerosis. Models were prompted with the clinical scenarios, and responses were scored for adherence to the CAR imaging recommendations. Results: Overall, 300 clinical scenarios were used to prompt each model. Adherence rates were 83.8% for GPT-4o, 94.0% for GPT-4o+RAG, 85.5% for o1, and 93.2% for o1+RAG. A Kruskal-Wallis test (H(3)=44.1, p<0.001) identified significant differences among models. Post-hoc comparisons showed RAG-enabled LLMs significantly outperformed standalone models (p<0.001). No significant differences were observed between GPT-4o and o1 without RAG (p=0.531), or between GPT-4o+RAG and o1+RAG (p=0.532). Conclusions: RAG integration significantly improved LLM adherence to Canadian neuroimaging guidelines, even when baseline models demonstrated moderate performance. Future work should validate and explore broader applications of RAG-enabled tools to advance evidence-based care.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.040 | 0.146 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.002 |
| Bibliometrics | 0.003 | 0.002 |
| Science and technology studies | 0.001 | 0.002 |
| Scholarly communication | 0.005 | 0.006 |
| Open science | 0.002 | 0.005 |
| Research integrity | 0.002 | 0.003 |
| Insufficient payload (model declined to judge) | 0.010 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".