Clinicians can ‘chat’ with medical records through new AI software, ChatEHR

In much the same way people can “chat” with large language models like GPT-4, Stanford Health Care clinicians can now interact with a patient’s medical records through an artificial intelligence-backed software called ChatEHR.

The technology, currently in a pilot stage, enables them to ask questions about a patient’s medical history, automatically summarize charts and perform other tasks. ChatEHR uses information from individual’s health records to provide its response.

“AI can augment the practice of physicians and other health care providers, but it’s not helpful unless it’s embedded in their workflow and the information the algorithm is using is in a medical context,” said Nigam Shah, MBBS, PhD, chief data science officer at Stanford Health Care, who led the team in developing the technology. “ChatEHR is secure; it’s pulling directly from relevant medical data; and it’s built into the electronic medical record system, making it easy and accurate for clinical use.”

The software has been in development since 2023, when a team of Stanford Medicine researchers led by Shah; Anurang Revri, vice president and chief enterprise architect for Stanford Health Care’s Technology and Digital Services; and others recognized the potential of large language models and were inspired to create something useful for clinicians.

“ChatEHR opens up a new way for clinicians to interact with electronic health records in a more streamlined and efficient manner, whether that’s asking for a summary of the entire chart or retrieving specific data points relevant to the patient’s care,” said Michael Pfeffer, MD, chief information and digital officer for Stanford Health Care and the School of Medicine, who helped lead the development and integration of the software. “This is a unique instance of integrating LLM capabilities directly into clinicians’ practice and workflow. We’re thrilled to bring this to the workforce at Stanford Health Care.”

Faster search, summary and information gathering

Currently, the software is accessible only to a small cohort of individuals at Stanford Hospital — 33 physicians, nurses, physician’s assistants and nurse practitioners — who are monitoring its performance, refining its accuracy and enhancing its utility.

“Making the electronic medical record more user friendly means physicians can spend less time scouring every nook and cranny of it for the information they need,” said Sneha Jain, MD, a clinical assistant professor of medicine who has been an early user of the technology. “ChatEHR can help them get that information up front so they can spend time on what matters — talking to patients and figuring out what’s going on.” 

When clinicians access the tool, they are greeted with: “Hi, 👋 I’m ChatEHR! Here to help you securely chat with the patient’s medical record.”

At that point, they can type in a slew of questions about the patient: Does this person have any allergies? What does their latest cholesterol test show? Have they had a colonoscopy? Were the results normal?

ChatEHR is not meant for medical advice, Shah said. The software is an information-gathering tool that can expedite the process and, ideally, save time. All decisions stay in health care experts’ hands.

Beyond a single search, ChatEHR can accelerate many of the time-consuming tasks that are part of a doctor’s everyday workload. Jonathan Chen, MD, PhD, a hospital physician and an assistant professor of medicine and of biomedical data sciences, noted that when a patient comes to the emergency room, the admitting doctor has to quickly figure out how to help them.

“It’s not just the chest pain they’re having in that moment that matters — it’s their whole story, what led up to this moment. All their prior history is relevant. What medications were they on, what side effects did they have, what surgeries took place and how did that affect them?” he said. “It’s a ton of work to go back and find all of that information during a time-sensitive case, so speeding up that process would be a big help.”

He added that ChatEHR could be helpful in some transfer cases. Patients who are transferred to Stanford Hospital for more advanced care generally arrive with a large packet of information, sometimes hundreds of pages long. “All of that medical history is crucial, but going in cold and sifting through that is a huge lift,” Chen said. “Having ChatEHR boil that down into a relevant summary would make that process smoother.” And, he said, it’s not just high-level summaries that ChatEHR provides, the physician can also ask probing follow-up questions to better understand the patient’s history.

The team is also building out something they call “automations,” or evaluative tasks based on a patient’s history and record. For example, the team has created an automation that can determine whether it’s appropriate to transfer a patient to the Stanford Medicine-affiliated Sequoia Hospital patient care unit, which offers more patient rooms. “That automated evaluation saves us the administrative burden of sifting through patient information and helps us quickly determine if a patient can be transferred, opening access to care here at Stanford Hospital,” Shah said. He and others are working on other automations, which would determine eligibility for hospice care, for example, or recommend additional attention post-surgery.

Continuing the rollout

Shah and the team will continue evaluating ChatEHR’s use cases using MedHELM, an open-source, flexible and cost-effective framework for real-world LLM evaluation in medicine. There are also other accuracy-ensuring features that are in development, such as citations that show clinicians where bits of information came from within the medical record.

As the technology develops, the goal is to open ChatEHR to all clinicians who look at patient charts. “We’re rolling this out in accordance with our responsible AI guidelines, not only ensuring accuracy and performance, but making sure we have the educational resources and technical support available to make ChatEHR usable and useful to our workforce,” Shah said.

Stanford’s Department of Medicine and the Center for Biomedical Informatics Research also supported the work.