Open Clinical Intelligence Lab
Accelerating clinical coding, healthcare knowledge systems and evidence generation through open science and AI-assisted research.
About OCIL
OCIL is an independent, open research lab founded by alumni of the University of Cambridge PACE Data Science, Machine Learning & AI programme. It started as our capstone project for NICE: a multi-agent system for clinical code discovery, now live at clinicalcodes.uk. We've kept building since — developing open methods, tools and datasets, and releasing them so the wider healthcare community can build on our work.
Projects
What we build and keep maintained in the open.
A LangGraph multi-agent pipeline that discovers and validates clinical codes across UK vocabularies (SNOMED CT, ICD-10, OPCS-4, dm+d, BNF). Six parallel retrievers, UMLS enrichment and per-code LLM scoring return a provenance-tracked codelist in seconds, behind an audited human-in-the-loop review gate. Built with NICE and the University of Cambridge.
An agentic pipeline for clinical-trial safety analysis. Inquest retrieves evidence from seven public sources, generates competing failure-mechanism hypotheses constrained to a published safety taxonomy, adversarially verifies every claim against cited evidence, and synthesises a calibrated, fully provenance-tracked triage memo for expert review.
Research
The areas we work across, from clinical AI down to the open infrastructure underneath it.
Where language models and ML actually help in the clinic: coding, retrieval and reasoning, built to be safe enough for real use.
The terminologies and ontologies that connect messy clinical data to how care is recorded in practice.
Careful, reproducible analysis of the routine data the health system already collects.
Turning vocabularies, guidelines and literature into a graph you can actually query.
Methods for finding and checking clinical evidence, with the working shown.
The shared tools and datasets that let other people reproduce what we do.
Publications
Open, citation-friendly outputs from the lab. Search by title, author or abstract.
C. J. Ramirez, A. Desalvo, D. Cage, A. Ramsawhook, Z. Li, I. Thanigaivelan
We present a LangGraph multi-agent pipeline that discovers and validates clinical codes across UK vocabularies (SNOMED CT, ICD-10, OPCS-4, dm+d and BNF). Six retrievers run in parallel over 53,000+ ingested codes, with UMLS synonym and hierarchy enrichment and per-code LLM scoring behind an audited human-in-the-loop review gate. Against 15 published OpenCodelists reference standards, mean F1 improved from 0.49 to 0.57 (p = 5.7×10⁻¹¹), with per-code rationale faithfulness of 83.5%.
C. J. Ramirez
Inquest is an agentic pipeline that analyses terminated clinical trials. It retrieves evidence from seven public sources, generates competing failure-mechanism hypotheses constrained to a published safety taxonomy, and subjects each atomic claim to adversarial verification against cited evidence before deterministic scorecard synthesis. On six verified terminated trials it achieved 5/6 top-1 mechanism matches with full citation traceability (1.00) and calibrated confidence.
Community
The people and institutions we build alongside, and the communities we want to work with.
These represent ecosystem and community categories and aspirational collaborators. They are not confirmed partnerships or endorsements unless explicitly stated.
Team
The people behind OCIL, working across clinical AI, data science and informatics.
Founder & Research Engineer
Research interests
Life Sciences Data Scientist
Research interests
Biomedical Data Scientist
Research interests
Data Science & Machine Learning
Research interests
Blog
Notes on what we're building, and what we're learning while we build it.
Our Cambridge capstone project for NICE turned days of manual codelist work into a query answered in under a minute. A short write-up of what we built, how it performs, and why manual curation needed fixing.
Most per-code scoring does not need a frontier model. We send the bulk of the work to Claude Haiku and keep Sonnet for the handful of steps that genuinely need it. Same quality, a fraction of the cost.
Inquest reads the public record on terminated trials, proposes the likely failure mechanisms, then tries to disprove each one against cited evidence before writing a short memo. It points reviewers at what the evidence supports, and stops there.
Contact
We collaborate with universities, healthcare organisations and open-source communities. Tell us about your project.