OpenMed provides an on-device clinical AI platform for healthcare organizations that need to extract, de-identify, and process medical data without sending patient information to external cloud services. The open-source, Apache-2.0 licensed software runs locally on hardware ranging from laptops to GPU servers, supporting 34 model-backed languages for PII detection and clinical entity extraction.
Funding
Funding not disclosed
Founders
Product
Problem
Healthcare organizations face a compliance dilemma when processing clinical notes: cloud-based medical AI APIs require sending patient data to vendor servers, while licensed enterprise NLP systems carry per-server subscription costs and limited language support. Existing research toolkits have largely stopped shipping updates, leaving teams without a current, cost-effective option for on-premises clinical text processing.
Solution
OpenMed delivers a local-first clinical AI platform that performs entity extraction, PII detection, and de-identification entirely on the user's own hardware, ensuring patient data never leaves the network. The Apache-2.0 licensed package installs via pip and provides a consistent API across multiple runtimes, including CPU-optimized ONNX builds, Apple Silicon MLX acceleration, mobile deployment through OpenMedKit, and browser execution via Transformers.js. A companion terminal-native agent runtime offers hybrid medical services for prior authorization appeals, coding, documentation, and care coordination, with visible plans, traces, and review checkpoints. The platform includes dedicated terminology services covering ICD-10, CPT, SNOMED CT, LOINC, RxNorm, HCC risk adjustment, and rare disease mappings, all accessible through local or protected endpoints.
Target Audience
Primary users are healthcare infrastructure builders, clinical informatics teams, and medical software developers who need compliant, on-premises NLP for clinical documentation, coding, and de-identification workflows.
Features
- One-call inference with structured outputs via `analyze_text(...)` for clinical entity extraction and PII detection
- Language-aware de-identification across 35 supported PII language codes with 34 model-backed languages
- BatchProcessor achieving up to 3.3× throughput on CPU and 2.2× on Apple Silicon MLX
- REST and gRPC service mode with built-in authentication and no-PHI logging
- Runs on iPhone, iPad, Android, Apple Silicon, NVIDIA GPUs, and in-browser via WebGPU
- CPU-optimized ONNX exports in fp32, fp16, and INT8 formats for hospital hardware without GPUs
- Hybrid medical service plane with dedicated endpoints for terminology, coding, drug lookup, and comorbidity scoring across 74,719 ICD-10-CM codes, 359,930 SNOMED CT concepts, and 19,598 CPT codes
- Weekly public release cadence with rerunnable benchmarks published on arXiv