OpenMed for Web

Run medical named-entity recognition and de-identification in browsers or Node.js with the same OpenMed models used by Python, Swift, and Android. Text is processed locally; loading a Hugging Face model may download its artifacts once, but inference does not send clinical text to a hosted API.

Install

For OpenMed models loaded through Transformers.js:

npm install openmed @huggingface/transformers

For direct ONNX Runtime Web sessions:

npm install openmed onnxruntime-web

Run an OpenMed ONNX Model

import { deidentify, loadOnnxModel } from "openmed";

const model = await loadOnnxModel("OpenMed/<model>-onnx-android");
const result = await deidentify(
  "Patient Alice Nguyen was seen in cardiology.",
  { pipeline: model },
);

console.log(result.deidentifiedText);
console.log(result.spans);

loadOnnxModel() selects the root INT8 model by default. Pass { variant: "fp32" } or { variant: "fp16" } when a different published variant is appropriate.

Local Browser Runtime

The lower-level ONNX Runtime Web loader keeps model and runtime assets on local paths and selects the strongest execution path available:

1. WebGPU
2. WebAssembly with SIMD and threads
3. Single-threaded WebAssembly

import {
  deidentify,
  loadOrtWebTokenClassificationPipeline,
} from "openmed";

const pipeline = await loadOrtWebTokenClassificationPipeline({
  modelPath: "/models/openmed/model.onnx",
  assetPath: "/models/openmed/onnxruntime/",
  tokenize: tokenizeClinicalNote,
  decode: decodeTokenClassificationOutputs,
});

const result = await deidentify(clinicalNote, {
  pipeline,
  detector: "ort-web",
});

Threaded WebAssembly requires cross-origin isolation. Serve the application with Cross-Origin-Opener-Policy: same-origin and Cross-Origin-Embedder-Policy: require-corp; otherwise OpenMed falls back to a single WebAssembly thread.

Privacy and Safety

• No telemetry is enabled by default.
• Local model and runtime paths reject remote URL schemes.
• Span records contain hashes and offsets rather than raw identifier text.
• OpenMed is not a medical device and must not autonomously make clinical decisions.

OpenMed is Apache-2.0 licensed. See the main repository for models, documentation, and the Python, Swift, and Android runtimes.