OpenMed 3.0 Released: Apache-2.0 Clinical AI Toolkit for Fully Local Patient Journey Aggregation

OpenMed 3.0, an Apache-2.0 licensed open-source clinical AI toolkit, has been released with a core focus on keeping patient data entirely on local hardware. The update introduces Patient Journey, a feature designed to aggregate and reconcile data from disparate sources such as clinic notes, FHIR feeds, HL7 messages, and lab CSVs into a unified patient timeline.

Key Features and Safety Constraints

The toolkit operates strictly offline, supporting 2,200+ open models on Hugging Face via Transformers, ONNX, GGUF (llama.cpp), WebGPU/WASM, and mobile devices. A central design principle is the refusal to guess or silently fallback to cloud APIs:

• Identity Resolution: A 99% identity match is not automatically merged; it requires human review to prevent cross-patient data contamination.
• Data Integrity: Conflicting diagnoses or corrected lab values (e.g., 6.8 to 4.4) are both retained and flagged for review rather than overwritten.
• Temporal Precision: Ambiguous dates (e.g., "January") are not auto-filled with default days to maintain data fidelity.
• Strict Local Execution: If a required local model is missing, the system refuses to process the request rather than calling a hosted API.

Release Status and Community Contribution

Version 3.0 is described as infrastructure plumbing for local models rather than a new clinical model checkpoint. The summarization pipeline is currently synthetic-only and not ready for real patient data. The release included 148 PRs from 10 contributors, with 46 from the community. The project is actively seeking contributors for version 3.1, which currently has 422 open issues and 30 good first issues available.