Developers can generate structured data and anonymized text from patient records locally
The process of automatically finding and removing sensitive information such as patient names, dates, and social security numbers from hospital records or clinical notes is completed on your computer or smartphone. This environment, which does not require sending data to cloud servers, is suitable for developers and researchers who prioritize medical data privacy. OpenMed operates in various languages including Python, Swift, and Kotlin, and allows you to run over 2,200 medical-specific models locally.
Data processing flow that ends locally
The core of OpenMed is the 'local first' principle. Once the necessary model files are downloaded locally, subsequent text analysis and personal information identification processes proceed without an external network connection. This fundamentally blocks the risk of patient data being transmitted over the internet. However, model downloads, remote provider integrations, and paths where telemetry features are enabled may use the network, so you should review the terms of use for each model and dataset. The core runtime performs extraction and anonymization locally once the required artifacts are ready.
Medical entity extraction and personal information de-identification
It automatically extracts medical-related information such as diseases, drugs, and anatomical structures from clinical text. For example, in the sentence 'Started imatinib for chronic myeloid leukemia treatment,' it identifies the disease and drug with high confidence respectively. Simultaneously, it can find and de-identify personal information (PII) in 18 categories, including names, dates of birth, phone numbers, and social security numbers, using various methods such as masking, replacement, hashing, and date shifting. The smart merge feature keeps dates or addresses as a single unit so they are not fragmented at the token level.
Support for various platforms and languages
On Apple Silicon-based Macs, accelerated inference is possible using the MLX runtime, and it can be integrated as a Swift package in iOS apps. In Android environments, it operates as a Kotlin library through ONNX Runtime Mobile, and in browsers, it can be executed with WebGPU acceleration using Transformers.js. Python requires version 3.10 or higher and supports various execution paths such as CPU, CUDA, and MLX. It also supports PII processing in 42 languages, of which 35 are backed by dedicated models. Russian uses the multilingual default model placeholder, while Bengali, Chinese, and Tamil have dedicated registry entries.
Items to check before deployment
Although the OpenMed SDK is released under the Apache-2.0 license, the licenses for models and datasets may differ, so you must check them before use. Regarding HIPAA compliance, this tool provides settings aligned with Safe Harbor standards, but using the SDK itself does not guarantee HIPAA compliance. Expert deployment review is still essential, and the deploying entity must verify the clinical suitability and privacy behavior of the models. Also, remember that the NER/NLI provider used in the 30-second demo is a synthetic test double, not an actual trained model or clinical validation.