A LOCAL-FIRST, PRIVACY-PRESERVING PLATFORM INTEGRATING THE FULL SYSTEMATIC REVIEW AND META-ANALYSIS PIPELINE WITH ON-DEVICE AI
Author(s)
Vitor P. Daldegan1, Lício A. Velloso, MD, PhD2.
1State University of Campinas (UNICAMP), Campinas, Brazil, 2University of Campinas, Campinas, Brazil.
1State University of Campinas (UNICAMP), Campinas, Brazil, 2University of Campinas, Campinas, Brazil.
OBJECTIVES: Systematic reviews (SR) and meta-analyses underpin health technology assessment (HTA), coverage, and reimbursement decisions, yet the workflow remains fragmented across single-purpose tools, and cloud-hosted AI raises data-governance and confidentiality concerns when evidence is sensitive or proprietary. We aimed to design and build an integrated, local-first platform covering the entire SR-to-publication pipeline, in which AI assistance runs entirely on the user’s machine and all methodological computation is deterministic and auditable.
METHODS: Compendium implements an 11-stage pipeline—dashboard, protocol, eligibility, search, import/deduplication, screening, full-text review, extraction, risk of bias, synthesis, and reporting—as a cross-platform Tauri/Rust desktop application with an embedded SQLite store, AI-assisted decisions via a provider-agnostic LLM layer, and automated PRISMA flow-diagram generation. Study data reside locally in SQLite; account metadata is cloud-stored. AI is local-first (Ollama/llama.cpp), with a human-in-the-loop trust boundary; research data stay on-device.
RESULTS: A complete 11-stage pipeline is implemented and packaged as a native Windows build (.exe, MSI, NSIS), with every stage navigable via a persistent workflow rail and command palette. Risk of bias uses three official instruments: a deterministic RoB 2 calculator (Sterne 2019), ROBINS-I v1, and ROBINS-E. Meta-analysis computes RR/OR/MD/SMD with DerSimonian-Laird random effects, heterogeneity (I², τ², Q, χ²), leave-one-out, and subgroup analysis; JAMA-style forest, funnel, and PRISMA 2020 flow diagrams export to SVG. Dual independent screening with Cohen’s κ (human-human and AI-human) and consensus resolution. No user-outcome or accuracy data are claimed — results describe implemented architecture and instruments, not clinical validation.
CONCLUSIONS: An integrated, local-first SR/meta-analysis platform is technically feasible without sending research data to external servers, combining official risk-of-bias instruments, standard meta-analytic methods, and a verifiable human-in-the-loop AI boundary in a single auditable workflow — designed to address confidentiality constraints common in health economics and outcomes research (HEOR)/HTA. A prospective validation cohort comparing platform-assisted and conventional review is the planned next step.
METHODS: Compendium implements an 11-stage pipeline—dashboard, protocol, eligibility, search, import/deduplication, screening, full-text review, extraction, risk of bias, synthesis, and reporting—as a cross-platform Tauri/Rust desktop application with an embedded SQLite store, AI-assisted decisions via a provider-agnostic LLM layer, and automated PRISMA flow-diagram generation. Study data reside locally in SQLite; account metadata is cloud-stored. AI is local-first (Ollama/llama.cpp), with a human-in-the-loop trust boundary; research data stay on-device.
RESULTS: A complete 11-stage pipeline is implemented and packaged as a native Windows build (.exe, MSI, NSIS), with every stage navigable via a persistent workflow rail and command palette. Risk of bias uses three official instruments: a deterministic RoB 2 calculator (Sterne 2019), ROBINS-I v1, and ROBINS-E. Meta-analysis computes RR/OR/MD/SMD with DerSimonian-Laird random effects, heterogeneity (I², τ², Q, χ²), leave-one-out, and subgroup analysis; JAMA-style forest, funnel, and PRISMA 2020 flow diagrams export to SVG. Dual independent screening with Cohen’s κ (human-human and AI-human) and consensus resolution. No user-outcome or accuracy data are claimed — results describe implemented architecture and instruments, not clinical validation.
CONCLUSIONS: An integrated, local-first SR/meta-analysis platform is technically feasible without sending research data to external servers, combining official risk-of-bias instruments, standard meta-analytic methods, and a verifiable human-in-the-loop AI boundary in a single auditable workflow — designed to address confidentiality constraints common in health economics and outcomes research (HEOR)/HTA. A prospective validation cohort comparing platform-assisted and conventional review is the planned next step.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
SA22
Topic
Medical Technologies, Methodological & Statistical Research, Study Approaches
Topic Subcategory
Literature Review & Synthesis, Meta-Analysis & Indirect Comparisons
Disease
No Additional Disease & Conditions/Specialized Treatment Areas