St Jude Releases MyGPT, an Offline Medical AI Tool

Researchers at St Jude Children’s Research Hospital have published MyGPT, open-source software that lets an organisation ask an AI model questions about its own documents without those documents leaving its systems. The tool is described in a paper in Nature Communications published on 24 October, and its code is free on GitHub under the GPL-3.0 licence.

MyGPT is a retrieval-augmented generation (RAG) system: it finds relevant passages in a library the user chooses, then has an open-source language model answer from them. The authors say it can query “biomedical literature, confidential clinical documents, and multilingual global health policies, while ensuring that sensitive data never leaves the secure environment”.

In brief

MyGPT runs open models locally or inside an institution’s own secure systems, according to the paper.
The authors list hardware for group or department use from a $1,839 Mac Mini to a $4,000 Nvidia DGX Spark.
Its “hallucination index” is the authors’ own score, which they describe as a guide for triage, not proof an answer is right.

What MyGPT does, according to the paper

The paper’s abstract says language models are held back in sensitive fields by “privacy concerns and risks of data leakage”, and that they “may hallucinate” when asked about information outside their training data. MyGPT tries to deal with both. It keeps the documents and the model in-house, and every answer shows the exact passages it drew on, so a user can check the answer against the source.

The software uses Ollama to run open models. The supplementary information lists families including Llama, Gemma, gpt-oss, DeepSeek-R1, Qwen3, Mistral and Vicuna. According to the README, the browser-based web app is the recommended way to use it, on a personal computer, a server or virtual machine, or Azure. A desktop app is available, though the macOS builds are not yet signed and notarised, so Macs may block them.

St Jude staff deployed MyGPT on virtual machines approved for protected health information, the paper’s acknowledgments say, and colleagues at St Jude Global helped evaluate its answers on health policies. The authors have also published the evaluation pipelines, covering six benchmark datasets including BioASQ, PubMedQA and kinase literature.

The hardware and cost figures the authors give

The authors present MyGPT as “a cost-effective, modular, open-source alternative to commercial LLM services”. Their supplementary information puts figures on that. Renting a cloud server with a single Nvidia GPU cost an average of $374.03 to $2,682.57 a month across the Google Cloud and Azure set-ups they priced. For a group or department, they list desktop hardware from a $1,839 Mac Mini with 64 GB of memory to a $4,000 Nvidia DGX Spark.

The team also tested the software on Apple M1 and M3 Macs, a St Jude server and an Nvidia Orin Nano, a small single-board computer. They say small models on such devices are “ideal for remote locations with limited or no internet access or for air-gapped deployments”.

The Nature page currently carries an unedited version of the paper with the abstract only, so TSN has not reported the paper’s accuracy results.

The hallucination index is a triage score, not a truth test

The name promises more than the score delivers, and the published peer review says so. In the transparent peer review file, one reviewer argued that the index measured how semantically close an answer was to the question and the sources, which “is not a proxy for factual correctness”. “The drug is safe” and “The drug is not safe” can look almost identical to such a measure.

The authors agreed with that point, added further scoring methods, including a machine-learning predictor of answer correctness, and now describe the index as “a guidance and triage metric, rather than a definitive measure of factual correctness”. That is the honest framing. A low score is a reason to look harder; a high score is not permission to stop checking. The cited passages are the real safeguard, and the reader still has to open them, as in TSN’s checklist for spotting made-up AI answers.

Keeping data on site is what opens the door for smaller hospitals

The more useful part of the release is the deployment model. For hospitals and health ministries whose records have to stay on their own systems, an open tool that runs on a desktop machine or a small offline device is an alternative to a commercial chatbot. Publishing the evaluation code also lets outsiders rerun the tests, as Microsoft did with its diagnosis benchmark.

The release does not show that MyGPT improves care. Like Google’s TB screening work, its value is likely to depend on how it performs where it is deployed. TSN tracks projects like this on its tech for good hub. For now, the test that counts is simple: whether a clinic with no internet connection and a $1,839 computer can get answers it can trace back to the page.