ChIRP: A ChatGPT Model for the NIH Intramural Community

BY ROBERT WAXMAN, CIT

Calling all humans: The NIH is piloting a generative AI service called ChIRP, short of Chatbox for the Intramural Research Program (IRP). Launched on January 7, this pilot program is powered by ChatGPT technology but caters solely to NIH staff and is confined to the NIH intranet.

The goal of the pilot is to perfect a ChatGPT service for NIH science and communication needs in the protective sphere of the NIH intranet to allow for the input of sensitive data, unpublished research results, and clinical research information. As such, CIT and its NIH partners who created the new platform need you to take ChIRP for a test drive—try it, push it, break it.

Register to participate in the ChIRP pilot at https://chirp.od.nih.gov. The pilot is open to all NIH staff.

ChIRP is many things

What is ChIRP? If you ask a commercial version of ChatGPT, you will learn that it is chromatin isolation by RNA purification. And that much is true. But we’re talking about a homegrown ChIRP, one not yet known by ChatGPT.

NIH ChIRP is a secure, generative AI environment that uses large language model (LLM) technologies. LLMs are virtual AI systems trained on vast amounts of data to “understand” language. Once trained, they can generate natural, humanlike responses and complete actions such as analyzing and summarizing content, coding, designing, translating, and brainstorming ideas.

ChIRP currently uses the GPT-4.o LLM developed by the OpenAI company and made available through the NIH STRIDES enterprise Azure cloud environment managed by CIT. ChIRP runs in an NIH environment, but no NIH internal information is integrated into the commercial models, and there is no data leakage outside of ChIRP or NIH.

This secure NIH environment allows staff to use sensitive data within ChIRP, like de-identified and anonymized clinical data, predecisional and draft policies, and nonpublic data, including scientific data and draft manuscripts. However, personally identifiable information is not permitted in ChIRP. Also, because the chatbot uses commercial models, it is not specifically trained or fine-tuned for NIH-specific or biomedical topics.

Overall, the pilot aims to create a secure environment for NIH staff to safely explore how generative AI technologies, such as LLMs and multimodal models (for example, those capable of processing images and other data types), can benefit NIH’s biomedical research. Feedback from pilot participants will inform future AI efforts at NIH.

ChIRP support is provided through the Clinical Research Informatics Strategic Planning Initiative (CRISPI), is funded by ODSS and OIR, and developed in collaboration with OD, CIT, NHLBI, and NIA.

Leaving the nest

The goal of the January launch event was to introduce ChIRP to the NIH community and help them understand what ChIRP is and to explain its potential uses, risks, and limitations. Event speakers also made sure attendees were aware of ethical concerns around using the tool such as the potential for amplifying biases, providing misinformation, exposing intellectual property, and plagiarism. Overall, presenters, including CIT Director Sean Mooney, were excited to share information and wanted attendees to leave the session with some basic, practical skills and encouraged everyone to get some hands-on time with ChIRP and share feedback.

During the presentation, the ChIRP team also provided a brief overview of the evolution of AI and LLMs, walked through the ChIRP interface, and offered tips on prompt engineering, the art of how to give the model instructions that yield the best results and reduce errors. A few quick tips: Start with simple prompts, be specific, and use clear instructions.

Presenters also demonstrated use cases for the tool, showing how the chatbot can summarize and mine information from uploaded text and documents, assume a role to provide domain-specific information, generate HTML code, and render graphics. One of the areas where ChIRP shines is analyzing long documents to provide summaries or pull specified information and organize it according to given instructions.

The need for ChIRP

ChIRP was born out of a need to help NIH researchers take advantage of AI technology and concerns around uploading confidential and sensitive information to public AI platforms, which are not necessarily private or secure. Mooney spoke at the launch event about the importance of exploring AI. “Timely access to advanced AI systems is really critical for advancing NIH research,” Mooney said. “ChIRP is a great example of how a pilot for an advanced technology could turn into an operational enterprise AI system to support the NIH.”

ChIRP also aligns with industry and government efforts exploring the responsible deployment and use of Al and Al-enabled technologies in the health and human services sector, including research and discovery, drug and device safety, health care delivery and financing, and public health.

CIT’s role in developing ChIRP

CIT played an important role in getting ChIRP up and running. CIT staff provided access to GPT-4.o models through a cloud-based STRIDES account and supported the development of ChIRP infrastructure in Microsoft’s Azure cloud environment.

CIT also played a key role in technical architecture and information security planning and supporting provisional authorization-to-operate, or ATO, efforts for the ChIRP pilot. Additionally, CIT provided communications support for the ChIRP launch, including messaging, video support, and design and graphic expertise for ChIRP-related materials.

A call to action

Feedback from pilot participants is crucial in helping to develop and refine the tool for NIH use and will inform future AI efforts at NIH. After you have used ChIRP, please provide your candid feedback through the ChIRP User Experience Survey. Feedback through the survey is anonymous.

To learn more, check out the recording of the ChIRP launch event and see the presentation slides on the event materials webpage.

If you have questions about the pilot, please reach out to the CRISPI team by email at CRISPI-LLM@od.nih.gov.