A globalized world demands ways to converge effort around a shared purpose. It makes me ask: what if we had a truly collaborative place to organize the health concepts the world relies on? What if we could speak a single clinical “code” and still be understood across systems, countries, and languages? Over the past year, I’ve helped move those questions closer to reality — translating 6k+ diagnoses into Portuguese, reviewing and fixing 15.5k+ writing/presentation issues, and supporting 30k+ mappings — while building the tools and QA workflows that make terminology work faster, safer, and more reproducible.
Me and the community
Hi — I’m Dr. Filipe, wrapping up a year-long OpenMRS Senior Fellowship. I’m a general practitioner in Brazil, working in Primary Care (Family and Community Medicine) and, at the same time, a software engineer for more than 15 years. In recent years, my focus has been on leveraging AI to improve processes and create sustainable new solutions — not as a “marketing feature,” but as a real tool to accelerate productivity, reduce repetitive manual work, and consolidate workflows that previously depended on constant human effort. I enjoy building systems that make operations safer, reproducible, and efficient: fewer lost spreadsheets, less “tribal knowledge,” more automation, traceability, and quality — less “machines’ work done by humans,” and more empathy in healthcare and in human work (that’s our motto at Orango).
When I first encountered the OpenMRS community, I was genuinely impressed by the project’s history, the maturity of its governance, and — above all — the engagement of people around the globe. For anyone who works in real-world healthcare, it’s rare to see such a strong combination of purpose, consistency, and open collaboration. As I worked toward joining this initiative, I saw a very concrete opportunity (through Dr. Andrew Kanter, who manages the CIEL terminology): to help improve the terminology ecosystem that supports a meaningful part of interoperability and data quality in OpenMRS implementations.
From Clinic to Concepts: Why CIEL Matters
As a physician in Brazil, using ICD-10 daily and experiencing its limitations in practice (ambiguity, lack of granularity, maintenance difficulties, translations, mapping rules), it made sense to “join forces” with the community to strengthen what already works well and accelerate what is still a bottleneck. CIEL (the Columbia International eHealth Lab’s open source concept dictionary at https://cielterminology.org) is a central piece of that story: it’s what many implementations rely on to maintain clinical consistency, and where small process improvements multiply into major global impact.
Between February 1, 2025 and January 21, 2026, I lived my first major cycle of structured work with OCL (Open Concept Lab at https://openconceptlab.org) and the CIEL/OpenMRS ecosystem. Instead of focusing on a single deliverable, the goal was to build and consolidate a set of tools and processes that simultaneously improved terminology quality, release speed and safety, the ability to detect issues before they reach production, and the scalability of tasks like translation, Quality Assurance (QA), mapping refactoring, and more.
From Lab to Product: CIEL Lab
CIEL Lab (first generation) was a platform “alongside OCL,” created to experiment with and operationalize tools tailored to the CIEL content — focused on optimizing terminology cleanup and maintenance workflows. During this period, it was used to detect and fix spelling issues and presentation inconsistencies, improve translation flows (French and Portuguese, including support for synonyms and short terms), verify mapping rules and incoherences, identify obsolete mappings, and support QA pages and routines that accelerate human review with more safety. Today, this first version is naturally approaching obsolescence, because several ideas matured and migrated into version 2.
CIEL Lab V2 consolidates what proved valuable in the first version, now with a more robust architecture, clearer abstractions for core workflows, a more consistent UX, and a foundation built to evolve. The goal was to reduce fragile couplings and make the tool more “product” and less “laboratory,” while preserving rapid iteration on improvements to CIEL operations alongside OCL. It also fills a practical gap today: the current OCL experience isn’t yet optimized for large-scale source concept management, so the Lab pages provide a safe place to build and validate new capabilities quickly — with the intent to move the best of them into OCL natively over time. The platform is integrated with the OCL API and is already progressing toward functioning as a full CIEL concept-management layer, using OCL as the source of truth, and helping remove long-standing friction between the terminology work and the management system (OCL), in parallel with TermBrowser v3. In other words: CIEL Lab has also become OCL Lab.
The Challenges of a New Language — ICD-11
One of the biggest real-world pain points is high-quality clinical coding, especially once we move beyond simple correspondences into cases that require post-coordination, extensions, and context. That’s why a significant part of this development cycle was the ICD-11 agent with a multi-step (agentic) approach, combining rules, heuristics, semantic (vector) search, reranking, and deterministic decision stages. The goal is not to “replace” specialists — nor even CIEL — but to reduce friction, accelerate triage within internal CIEL processes, and improve consistency when we apply terminology rules at scale, especially in scenarios where human cost becomes the bottleneck.
It’s important for CIEL users because ICD-11 is not a simple “lift-and-shift” upgrade from ICD-10. Its structure is fundamentally different: instead of relying only on a fixed, pre-coordinated list of diagnoses, ICD-11 enables post-coordination — combining a base code with extensions and contextual qualifiers. That unlocks much richer clinical expressiveness, but it also makes adoption and migration significantly more complex for countries moving from legacy coding practices and systems. And this is a global shift that will have to happen: for example, Brazil has set a target to transition to ICD-11 by January 2027, precisely because it requires system updates, training, and a carefully managed rollout. In practice, we can see this complexity in how many CIEL concepts map to post-coordinated ICD-11 expressions (i.e., more than one ICD-11 code is needed to faithfully represent a single clinical concept). So far, 8,531 of 25,758 mapped CIEL concepts (~33.1%) are post-coordinated when considering all map types; and even among SAME-AS mappings, 3,854 of 9,193 concepts (~41.9%) require post-coordination. These proportions reflect the real-world lift of ICD-11 adoption — and why keeping an interface terminology like CIEL central to clinical workflows is so valuable for a seamless migration without migraine. This is particularly true for all the existing information system mappings from local dictionaries to application forms, reports and analytics. The interface terminology (CIEL) allows for seamless mapping to ICD-11 while future-proofing any additional changes to the administrative terminology.
Designing for Scale: Memory, Safety Nets, and Controlled Change
In the same spirit of integrating modern tooling into the ecosystem, I also worked on a client for the newly created OCL MCP, with the goal of facilitating integrations with agentic flows and enabling external solutions to “talk” to OCL in a more productive way, standardized by the MCP Server tooling (developed by Jonathan Payne) and made available for non-technical users to test through an intuitive interface. For me, this kind of bridge is essential: it connects what the community already has solidly (governance, content, APIs) to new forms of interaction and automation. More on the OCL MCP coming soon!
Another important part of my work was structuring collections and sources following OCL standards, helping make content more reproducible and easier to maintain. This includes initiatives such as an ICD-10 French source, a hierarchical WHO ATC, and work with Monkeypox and COVID collections within CIEL. This type of work may look “invisible,” but it’s what maintains consistency and scalability when many people and organizations depend on the same ecosystem — and it must soon be replicated for other terminologies as well. Other external terminology processes were carried out with ICD-11, IMO, SNOMED, and RxNorm to support integration into CIEL workflows.
For mapping to these external terminologies, I also invested in concept vectorization and semantic search mechanisms to accelerate tasks that, in practice, depend on “finding the best candidate” at low operational cost. This kind of infrastructure becomes the basis for code suggestion, candidate retrieval, faster review, and better internal tooling. When done well, it reduces dependency on repeated manual attempts and increases the average quality of the process.
A critical point in terminologies is what happens after a release: new codes enter, others become obsolete, relationships change — and without a verification process, problems only show up after something breaks in real-world usage. We also face the challenge of reviewing terminologies with thousands or millions of concepts, which don’t always provide a clear changelog of what was altered since previous versions. Because of that, I built automations to verify obsolete concepts and mappings when new releases arrive from sources such as WHO ATC, SNOMED, and RxNorm — helping detect impact, replacement needs, and post-update inconsistencies before they become downstream problems. This makes the path to new CIEL versions much smoother.
In addition, I worked on automating mapping refactoring based on externally curated data — especially suggestions from IMO Health. The key here was applying changes in bulk with control, validating inconsistencies, generating diffs/audit trails, and creating “blacklist” mechanisms when needed. This balance between scale and safety is essential: automation without validation becomes risk; automation with traceability becomes a multiplier of impact.
I Work with Numbers
In terms of cycle metrics, a few numbers from CIEL help convey scale: 6k+ diagnoses translated into Portuguese, 15.5k+ writing/spelling correction items analyzed and fixed, dozens of clinical concepts manually corrected when automatic rules weren’t enough, and more than 30k mappings performed with workflows optimized by the process improvements established. To me, these numbers aren’t “vanity”; they represent real operational workload that, once turned into tooling and process, stops depending solely on heroics and starts depending on engineering.
Past Lessons, Future Directions
What I learned most through this process is that “good” terminology isn’t just content: it’s pipeline, QA, traceability, and governance. AI works best when coupled with rules, well-prepared data, and clear validation — not as a black box. And above all, an open and well-governed community accelerates everything: when there is real feedback, review, and alignment, the end result improves faster and with more confidence. Of course, we can’t forget that all of this is enabled not only by tools, but primarily by the people who use them — a cohesive team, capable people, and an excellent environment to work in certainly made the process lighter, more fun, and possible to deliver what CIEL/OCL needs.
The natural next steps are to consolidate stability and transition of the tools (especially CIEL Lab V2), strengthen integrations with OCL, and keep advancing what truly changes the game: more reproducible workflows with strong observability, safer automation, greater ability to scale with quality, and tools that bring clinical practice and technology closer together. If you work with OpenMRS, interoperability, terminologies (ICD, SNOMED, LOINC, RxNorm, ATC), or you’re interested in applying AI pragmatically in this space, I’d love to connect — this year was only the beginning.
In closing, I’d like to record my gratitude for Andrew Kanter (CIEL | OpenMRS), Jonathan Payne (Open Concept Lab), Joe Amlung (Open Concept Lab | Regenstrief), Grace Potma (OpenMRS), Sunny Aggarwal (Open Concept Lab), and Erica Kigotho (OpenMRS) — for their welcome, patience, charisma, proficiency, and steady availability throughout this work.
MD | Software (AI) Engineer
Leave a Reply