LUMINA: Agentic AI Framework for Systematic Review Automation

---

📌 Overview

LUMINA (Learning-based Unified Multi-agent for INtegrated Article screening) is an agentic AI framework that leverages large language models (LLMs) to transform how systematic reviews and meta-analyses (SRMAs) are conducted.

Manual screening in SRMAs is notoriously time-consuming, error-prone, and inconsistent. Our framework introduces multi-agent collaboration, Chain-of-Thought reasoning, and peer-review style auditing to achieve human-level sensitivity with higher efficiency and reproducibility.

Published in NEJM AI (2025), LUMINA demonstrates how LLMs can augment medical evidence synthesis with transparency, scalability, and cost efficiency.

---

🚀 Key Features

• Multi-Agent Architecture

  • Classifier Agent – triages citations (Potentially Relevant, Likely Irrelevant, Uncertain) with bias toward sensitivity.
  • Detailed Screening Agent – applies structured Chain-of-Thought (CoT) prompts aligned with the PICOS framework (Population, Intervention, Comparison, Outcome, Study design).
  • Reviewer Agent – audits decisions, ensures consistency, and invokes improvement when needed.
  • Improvement Agent – performs self-correction and refinement when conflicts or errors are detected.

• Research-Grade Performance

  • Sensitivity: 98.2% (SD 2.7%)
  • Specificity: 87.9% (SD 8.4%)
  • Outperformed state-of-the-art LLM screening baselines in 4 out of 15 SRMAs.

• Transparent + Scalable

  • Decisions explained via structured reasoning.
  • Can be scaled to thousands of citations per review.

• Cost-Efficient

  • Average screening cost of ~$0.07 per 10 articles.
  • Processing time ~6.7 minutes for 10 citations.

---

🧠 Why LLMs?

Traditional machine learning models struggle with nuanced reasoning in biomedical literature.
LLMs (GPT-4, GPT-4o, Claude, etc.) allow:

• Semantic understanding of complex clinical abstracts.
• Structured reasoning via Chain-of-Thought prompts.
• Self-reflection and correction when combined with agentic orchestration.

LUMINA harnesses these capabilities in a multi-agent loop that mimics the human peer-review process.

---

🏗️ System Architecture

🧮 Classifier Agent

• Role: First pass triage of every citation.
• Outputs: Potentially Relevant | Uncertain | Likely Irrelevant.
• Events Emitted: triage.event.

🕵️ Reviewer Agent

• Role: Acts as the first gatekeeper after each classifier or screener decision.
• Decisions:
  • agree? → whether the decision is logically consistent.
  • include? → whether to forward to the next stage.
• Events Emitted: review.event.
• Notes: Mirrors the two diamond decision points in the system diagram.

🔄 Improver Agent

• Role: Self-correction loop.
• Triggered When: agree? = false at any gate.
• Action: Generates an amended rationale/label and re-routes output back to the Reviewer.

📑 Detailed Screening Agent

• Role: Secondary screening for items marked as Potentially Relevant or Uncertain and include? = true at the first gate.
• Method: Produces PICOS-grounded Chain-of-Thought reasoning plus a proposed final label.
• Events Emitted: screen.event.

✅ Decision Sink

• Include Set: Final accepted citations (green checklist).
• Discard: Final rejected citations (yellow oval).
• Terminal State: Every citation must end in either Include or Discard.

📌 Invariant: Every decision made by Classifier or Detailed Screening Agent must pass through the Reviewer. Any disagreement is resolved through the Improver loop before advancing.

---

flowchart LR
  A[Classifier] --> R1[Reviewer]

  %% First review cycle
  R1 -->|Disagree| I1[Improver] --> R1
  R1 -->|Include: Potentially Relevant / Uncertain| S[Detailed Screening]
  R1 -->|Include: Likely Irrelevant| D[Discard]
  R1 -->|Exclude| D

  %% Detailed screening cycle
  S --> R2[Reviewer]
  R2 -->|Disagree| I2[Improver] --> R2
  R2 -->|Include| INC[(Include)]
  R2 -->|Exclude| D

Loading