Sovereign Systematic Review & Meta-Analysis Engine

A zero-egress clinical literature extraction pipeline that uses Apertus 1.5's 262k context and deterministic bounding-box coordinate locks to automate medical meta-analyses without LLM hallucinations.

Comment

Inspiration

Clinical researchers and healthcare institutions spend months manually extracting PICO metrics (Population, Intervention, Comparator, Outcome) from hundreds of dense clinical trial PDFs for systematic reviews and meta-analyses.

Current cloud-based LLMs fail here for two critical reasons:

  1. Data Sovereignty Violations: Embargoed clinical trial data, pre-print studies, and proprietary hospital documents cannot legally leave internal servers to hit external commercial APIs due to strict GDPR, HIPAA, and data protection laws.
  2. The "Vegas Veto" (Hallucinations): If a generalist LLM hallucinates an N-value sample size (e.g., reporting 40 instead of 400) or misquotes an odds ratio, the entire meta-analysis is poisoned.

We built the Sovereign Systematic Review & Meta-Analysis Engine to prove that Switzerland's sovereign open LLM, Apertus 1.5, combined with deterministic physical coordinate verification, can fully automate literature extraction while remaining 100% auditable and air-gapped.

What it does

  • Air-Gapped Ingestion: Ingests complex multi-page clinical trial PDFs locally without external network calls or data egress.
  • Extreme-Context PICO Structuring: Leverages Apertus 1.5's massive 262k context window and native <|inner_prefix|> Thinking Mode to extract structured PRISMA metrics, study design, patient sample sizes, and Cochrane Risk-of-Bias 2 assessments.
  • Deterministic Bounding-Box Coordinate Lock: Solves the hallucination bottleneck. Every metric extracted by Apertus must return an exact verbatim quote from the text. A native Python layer (PyMuPDF) locates the exact page number and physical [x0, y0, x1, y1] bounding box coordinates in the PDF, highlighting the source text in bright yellow.
  • Automated Deliverables: Generates both a verified, PRISMA-compliant CSV matrix for immediate statistical analysis and visual Audit PDFs for human-in-the-loop verification.

How we built it

  • Sovereign LLM Engine: Powered by Apertus 1.5 (utilizing both hosted API endpoints and local 4-bit NF4 quantized weights via andreasmartin/apertus-v1.5-8b-text for air-gapped GPU deployment).
  • PDF Vectorisation & Verification: Built with PyMuPDF (pymupdf) to parse multi-page layouts and execute deterministic bounding-box annotation passes without file corruption.
  • Structured Schema Enforcement: Implemented strict Pydantic and JSON schema validation with thinking-trace separation.
  • Deployment & UX: Packaged as an on-premises Docker container launched via a simple make run command, featuring an interactive Gradio web dashboard.

Challenges we ran into

  • Incremental PDF Writing Collisions: PyMuPDF throws ValueError: save to original must be incremental when attempting to apply multiple highlight annotations to an open document. We engineered a single-pass multi-quote verification engine that buffers document mutations cleanly.
  • Apertus Thinking-Mode Extraction: Parsing structured JSON when the model generates <|inner_prefix|> reasoning traces required building custom regex sanitisers to cleanly decouple internal reasoning tokens from the final JSON payload.
  • Zero-Egress Containerization: Designing a lightweight Docker container that satisfies the Hack Apertus Track 2B mandatory target architecture (on-premises / air-gapped/sovereign cloud) while handling complex C-bindings for PDF rendering.

Accomplishments that we're proud of

  • Defeating the "Vegas Veto": Building a system where every single number in the final meta-analysis matrix is physically tethered to an auditable coordinate on the source PDF page.
  • Strict Architecture Compliance: Adhering strictly to the Hack Apertus Track 2B mandate: fully functional on-premise deployment via make run in Docker with a sub-100MB footprint.
  • Real-World Impact: Delivering a high-utility research tool that directly solves the literature review bottleneck for medical researchers, universities, and pharmaceutical teams.

What we learned

  • Sovereign, open-weight models like Apertus 1.5 with large context windows can match proprietary models on domain-specific document structuring when combined with rigorous schema enforcement.
  • Trust in medical AI cannot rely on probabilistic model confidence; it demands deterministic physical anchors back to source data.

What's next for Sovereign Systematic Review & Meta-Analysis Engine

  • Cross-Lingual Extraction: Expanding the pipeline to utilise Apertus's 1,811-language training to extract and synthesise non-English clinical trials (French, German, Spanish, Chinese) to eliminate language bias in global meta-analyses.
  • Interactive Coordinate Viewer: Building a dual-pane web UI where clicking any cell in the PRISMA table instantly scrolls the PDF viewer directly to the highlighted bounding box.
  • Automated Statistical Synthesis: Integrating an R/Python meta-regression module to generate forest plots and funnel plots directly from the extracted CSV matrix.

Built With

Share this project:

Updates

Submission history