AI‑Assisted Delphi for RNA‑Based Medicines: A Controlled Comparison of Human and Automated Revision (Preprint)
Bibliographic record
Abstract
BACKGROUND: The Delphi method is widely used to derive expert consensus on complex clinical problems, yet it is slow and resource intensive. Recent advances in large language models and retrieval‑augmented generation (RAG) offer the possibility of accelerating consensus while maintaining methodological rigor. Large language models can retrieve and summarize evidence, but they frequently hallucinate and cannot reliably cite sources. At the same time, RNA‑based drugs and messenger RNA vaccines are rapidly moving from concept to clinic, generating a pressing need for timely, evidence‑based consensus on regulatory, manufacturing, and clinical issues. OBJECTIVE: We evaluated whether a modular, RAG‑enabled, multi‑agent artificial intelligence (AI) pipeline could replicate the post-round 1 behavior of a human reviewer in a Delphi study. The primary objective was to determine whether AI‑assisted statement revision could rescue a greater proportion of subthreshold statements and achieve a consensus comparable to that obtained through human revision by round 2. METHODS: A parallel, 2‑arm Delphi study was conducted on 28 statements about RNA medicines. In total, 50 international panelists (clinicians, researchers, and patient representatives) were randomized into human (arm A) and AI‑assisted (arm B) groups. After round 1, statements below the 75% agreement threshold were revised either manually by human reviewers or by an AI pipeline comprising the following software agents: (1) ReferenceDetector, to identify external citations; (2) Summarizer, to produce structured summaries of supporting PDFs; (3) a hybrid RAG module that combined dense and sparse retrieval with cross‑encoder reranking; and (4) Refiner, which generated revised statements, reasoning logs, and explicit citations. Two human reviewers with expertise in Delphi methodology and literature review who had contributed to statement development verified retrieved citations and approved or amended revisions. Agreement rates and vote distributions were compared across arms. RESULTS: Arm A reached consensus on 71.4% (20/28) of the statements in round 1, whereas arm B reached consensus on 46.4% (13/28). After revision, consensus increased to 92.9% (26/28) of the statements in arm A and 85.7% (24/28) in arm B. The AI arm exhibited a larger mean improvement (absolute difference between rounds 1 and 2=39.3 percentage points) because more statements were initially below the threshold. Nonetheless, the absolute difference between arms after round 2 was modest (7.2 percentage points). AI‑assisted revisions were particularly effective for statements far below the threshold, but both arms failed to rescue 2 to 3 statements owing to substantive disagreements. CONCLUSIONS: A modular, citation‑anchored AI pipeline can closely approximate human performance in Delphi consensus procedures while substantially reducing manual workload. When paired with human oversight, AI assistance accelerated revision and closed most of the performance gap by the second round. Adoption of AI‑assisted workflows could accelerate consensus development on emerging technologies such as RNA therapeutics provided that transparency, rigorous retrieval, and human review are maintained.
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 | 0.107 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.002 | 0.001 |
| Science and technology studies | 0.002 | 0.001 |
| Scholarly communication | 0.002 | 0.001 |
| Open science | 0.001 | 0.003 |
| Research integrity | 0.001 | 0.001 |
| Insufficient payload (model declined to judge) | 0.010 | 0.002 |
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".