Open-source healthcare AI projects worth knowing in 2026
Ten open projects for clinical NLP, de-identification, medical imaging, synthetic data, FHIR and medical language models, and what each one is good for.
- Global
Key takeaways
- De-identification: OpenMed (Apache-2.0, 35 PII languages) and Microsoft Presidio (MIT) both run on your own hardware.
- Synthea generates realistic synthetic patient records, so you can build and test without touching real data.
- HAPI FHIR is the most widely used open-source FHIR server and library for Java.
- Open-weight medical models such as MedGemma ship under their own terms. Read them before you build a product on top.
Which open-source projects matter for health AI?
| Project | What it does | Licence | Good for |
|---|---|---|---|
| OpenMed | Clinical NER and de-identification, 2,200+ models | Apache-2.0 | Removing identifiers on-device, in 35 languages |
| Microsoft Presidio | PII detection and anonymisation framework | MIT | Pipelines that mix rules, regex and models |
| medspaCy | Clinical NLP on top of spaCy | MIT | Rule-based clinical text processing, negation and context |
| MedCAT | Concept extraction and linking to SNOMED CT and UMLS | Check current terms | Turning notes into coded concepts |
| Apache cTAKES | Clinical text analysis from Mayo Clinic, on the JVM | Apache-2.0 | Legacy hospital NLP pipelines |
| Project MONAI | Deep learning framework for medical imaging | Apache-2.0 | Training and deploying imaging models |
| Synthea | Synthetic patient generator | Apache-2.0 | Test data with no privacy risk |
| HAPI FHIR | FHIR server and Java library | Apache-2.0 | Interoperability with EHRs |
| MedGemma | Open-weight medical text and image models from Google | Health AI Developer Foundations terms | Medical text and image understanding |
Where should a new health AI team start?
Start with data safety, not models. Put de-identification (OpenMed or Presidio) between your users and anything that stores or learns from their messages. Use Synthea to build and test flows without real patients. Add HAPI FHIR when you need to talk to clinical systems. Only then pick a model, and pick it with an evaluation set, not a leaderboard.
What should you check before adopting a project?
- Licence: Apache-2.0 and MIT are permissive. Some projects have moved to source-available licences, and many models ship under custom terms.
- Release cadence: a project that has not shipped in a year will not fix the bug you find.
- Languages: most clinical NLP is English-first. Test every language your users write in.
- Evaluation: can you reproduce the published numbers? Projects that publish weights, recipes and evaluation scripts let you check.
Frequently asked questions
Is it safe to use open-source models for patient-facing health AI?
The licence and the code are only the start. Safety depends on how you test the model on your users, languages and risky scenarios, and on the guardrails around it. Evaluate before launch and on every update.
What is the difference between open source and open weights?
Open source means the code is under a licence such as Apache-2.0 or MIT that allows use, change and redistribution. Open weights means you can download the model parameters, but the terms may restrict uses. Many medical models are open weights, not open source.
Sources
- OpenMedOpenMed · 2026-11
- PresidioMicrosoft on GitHub
- medspaCyGitHub
- MedCATCogStack on GitHub
- Apache cTAKESApache Software Foundation
- Project MONAIMONAI Consortium
- SyntheaThe MITRE Corporation
- HAPI FHIRSmile Digital Health
- MedGemmaGoogle Health AI Developer Foundations · 2025-05
Found something out of date? Write to contact@brandlabs.app and we will fix it.