Work

A Prompt-Driven, Human-Supervised AI Research Infrastructure for Population-Scale Epidemiological Studies: Development and Validation

Year: 2026
Type: preprint
Abstract: Background: Clinical research infrastructure imposes prohibitive costs and institutional dependencies that systematically exclude solo investigators and under-resourced research groups. Existing AI tools assist individual research tasks but do not provide an integrated, end-to-end research environment paired with rigorous human oversight. We describe the design, implementation, and validation of Fourier Health Analytics Lab (FHAL), a human-supervised, agent-assisted AI research infrastructure built and operated by a single investigator through natural language prompts to Claude Code — no code was written directly by the investigator — with continuous investigator review and approval of all generated outputs. Methods: FHAL was constructed across five sequential phases using prompt-driven development through Claude Code (Anthropic). The system integrates a persistent SQLite state database, a command-line interface, automated scheduling, a multi-source literature monitor spanning PubMed and Semantic Scholar, a ChromaDB vector memory layer, and specialised analysis and writing agents. Validation was performed using an epidemiological study with a pre-specified analytical plan of consumer cosmetic chemical exposures and systemic inflammation, drawing on NHANES 2005-2010 data (N=4,864; complete cases across seven exposure biomarkers). Results: FHAL successfully executed an end-to-end research pipeline from literature ingestion to submission-ready manuscript production. The literature monitor ingested 1,467 papers on first run across 17 search lanes, classified all papers via Claude API with PICO extraction and relevance scoring, and fired 34 Telegram alerts for high-relevance findings. The analysis agent processed 4,864 NHANES participants across three survey cycles (from a merged adult dataset of 18,161 adults aged ≥18) using survey-weighted regression, false discovery rate correction, weighted quantile sum mixture analysis, multiple imputation under Rubin's rules, and sex-stratified sensitivity analyses. Key findings included significant positive associations between BPA and log-CRP (FDR-corrected p<0.05) and a substantial WQS mixture effect dominated by monocyclohexyl phthalate. The writing agent produced a complete JAMA Internal Medicine manuscript package including draft, simulated peer review, cover letter, and policy brief. Total elapsed construction time was approximately 18 hours of investigator-directed prompt sessions, with all generated code reviewed by the investigator before deployment. Conclusions: Human-supervised, agent-assisted research infrastructure is feasible, reproducible, and scientifically rigorous when paired with continuous investigator oversight. FHAL demonstrates that a solo investigator can build and operate a research platform capable of executing population-scale epidemiological studies, monitoring global literature, and producing publication-ready manuscripts without authoring code line-by-line, while retaining full scientific and ethical responsibility for every output. This approach has significant implications for democratising research capacity in under-resourced settings and for accelerating the pace of evidence generation in global health.
Language: English
Field: Medicine
Open Access status: green
OpenAlex ID: W7164167347
Cites: 15
Cited by:
Related to: 10