AI tool assists doctors in sharing lab results

Stanford Medicine physicians have a new artificial intelligence tool to assist them when they message patients about test results. The technology drafts an interpretation of clinical test and lab results and explains them in a message using plain language, which a physician then reviews and approves.

The technology adds to a growing number of AI-based tools poised to help physicians spend less time on administrative tasks and more time on more meaningful work, such as interacting with patients.

“Artificial intelligence has tremendous promise to enhance the experience of both patients and clinicians in the health care setting — and this tool is one of many ways that we are unlocking that potential,” said David Entwistle, president and CEO of Stanford Health Care. “At Stanford Medicine, we are proud to be at the forefront of implementing responsible AI in clinical care, with a focus on advancing and empowering better health for all.”

The tool, built in-house, leverages Anthropic’s Claude 3.5 Sonnet LLM through Amazon Bedrock, an Amazon Web Services service for building, deploying and scaling generative AI applications. The tool follows a model similar to the AI-powered draft message generator that creates responses to patient inbox messages for health care providers.

“It’s really about continuing to give back to the patient-provider relationship,” said Michael Pfeffer, MD, chief information officer at Stanford Health Care and the Stanford School of Medicine. “We hope this helps make time for that through reduced administrative tasks and by facilitating better communication and faster communication.”

Testing the tool

When a physician orders a test for a patient, be it blood work, an X-ray, a biopsy or something else, the results appear as medical data without interpretation. By law, those results must be shared with the patient as soon as they’re ready, but the descriptions are often highly technical and can be challenging to understand without a medical degree. That means physicians are tasked with translating the results for their patients.

“Millions of lab tests are ordered annually by physicians for their patients,” said Aditya Bhasin, vice president of software design and development at Stanford Health Care. “By generating draft responses for our physicians, we not only assist them with that workload, but also provide timely, comprehensive comments to help patients understand their specific results.”

In a pilot test of the technology, 10 primary care physicians used the tool for a month, and with their feedback incorporated, the team ran the second pilot among a cohort of 24 physicians for an additional two months. The drafts aren’t sent directly to the patient after the tool creates them — each one is reviewed and, if necessary, edited, before the physician sends the note.

“As a clinician, I love that I don’t have to start with a blank page and the draft is in language that’s understandable for patients,” said Christopher Sharp, MD, chief medical information officer at Stanford Medicine. “I’ve had patients say to me, ‘Dr. Sharp, you always write a comment on my result, and it makes me feel so much better.’ It takes effort and time to create those notes in a clear and empathic way, and I think this tool will make it easier and more efficient to provide those interpretations, which are so important to our patients.”

The goal is not for the tool to replace the physician’s messaging interaction with the patient, but rather provide the physician with a draft that’s either ready or close to ready to send, Pfeffer said. Of course, there’s no obligation for the physician to use it.

Guided by principles of the RAISE Health Initiative, which focuses on the development and implementation of AI technologies at Stanford Medicine with responsibility and safety at the forefront, and the FURM (fair, useful, reliable models) assessment, Bhasin and his teams evaluated multiple models to find the best fit for Stanford Health Care and to ensure the tool did not introduce unintended biases or kinks in clinical workflows.

The team received positive anecdotal feedback during the initial introduction of the software.

“We’ve gotten great responses from physicians so far,” Bhasin said. “Our physicians have shared that they appreciate the drafts’ concise and accurate nature, that they’re personalized for the patients, and that the messages are reassuring when communicating normal results.”

A boost for primary care and beyond

For now, primary care physicians across Stanford Health Care are benefiting from the tool; the plan is to roll it out to specialists this year. The team also plans to measure the success of the tool, analyzing data that shows time saved and how often the AI-generated drafts are used, among other measurements.

“We are thrilled to have the opportunity to provide these technologies to our physicians and study them and learn from them,” Pfeffer said. “It’s still fairly new, and it’s going to get better and more valuable as we develop it and incorporate feedback from our physicians.”