Main content

How Generative AI can streamline medical workflows

George Lawton Profile picture for user George Lawton March 3, 2023
Is healthcare the use case for ChatGPT?

healthcare, people, technology and medicine concept - close up of doctor in white coat with stethoscope and tablet pc computer over blue background with charts © Syda Productions - Shutterstock
(© Syda Productions - Shutterstock)

ChatGPT has driven an incredible level of hype about generative AI techniques. But applying these techniques to business will require some care, particularly in safety and regulatory-conscious industries like healthcare.

Hospitals have zero tolerance for the kinds of misinformation and hallucinations that have cropped up with ChatGPT, Google Bard, and Microsoft Chat. At the same time, early adopters of new generative AI tools see tremendous upside if they get it right.

One such early adopter is The University of Kansas Health System, which recently launched a pilot program with Abridge to streamline its clinical documentation workflow that could eventually reach 1,500 practicing physicians. The hope is that generative AI could automatically summarize patient encounters and dramatically reduce doctors’ work outside the exam room. It will also keep patients in the loop about what doctors said and what they might have missed in the exam room.

Dr. Gregory Ator, Chief Medical Information Officer and Head and Neck Surgeon at The University of Kansas Health System, hopes the new system can bring joy back into medicine, adding: 

We doctors did not go to medical school to do clerical transcription from talking to patients.

He has been leading efforts to digitize medical records since 2003 which both improved access to information and complicated workflow for doctors. Ator explains:

Part of the problem is there are so many consumers of information. About ten different sets of people want to look at doctor’s notes. When we were doing medicine on index cards, those other folks were not as prominent as they are now. We are trying to create a document with multiple consumers, and there is a lot of friction. After all that time creating the records, the second thing doctors want to spend time on is hunting and gathering information.

From transcription to translation

Ator later led the rollout of Dragon’s transcription tech to help streamline the documentation process. The current approach uses Dragon combined with other tools to structure part of the data. They also use about eight other tools to automate parts of workflows, such as repeating frequent blocks of text. But even with automated transcription, physicians have been spending an average of 130 minutes daily on documentation outside of medical hours.

The big difference with these new generative AI techniques is that they can automate translating and summarizing information for different audiences based on the doctor’s conversation with the patient. This reduces the need for a secondary dictation process after the visit. Ator says:

I am not interested in a courtroom-like transcription. I am interested in an accurate summary in a human consumable format that reflects what happened in the room, does not leave any details out, and does not add anything that did not happen.

Structuring the data is complicated for us, but it is also essential for us to be better doctors. Every doctor hopes that technology can make them better.

Simplifying processes

Another important aspect is that the documentation happens in line with the work rather than as a secondary process after the doctor may have forgotten details. This is critical for doctors who see multiple patients with similar problems. It also means that the doctor can go home at the end of the day after seeing patients. Ator explains:

Real-time is important. If an orthopedic surgeon sees ten to twenty patients daily, many of the questions and answers they give patients are similar. So doctors sometimes rush through the clinic and document at the end of the day. But then remembering who said what is important.

When people do that delayed documentation, the time to produce that note goes up, and the quality goes down because of obvious human factors and people trying to do that during family time.

This new approach also has the opportunity to include the patient in the documentation process as a participant rather than someone who is talked about after the fact. The new tools also provide a consumer portal for delivering summarized patient notes. Patients can look at the summary and then play back what the doctor said in the exam room when they do not understand something. This allows the patient to go over what was said and the words that went with it in great detail.

The trust factor

The early goal is an accurate translation of voice into a distilled record of what happened. This will make it easier to structure the data to research how doctors have treated similar patients and their outcomes both within the hospital and nationally. Ator also wants to demonstrate the new system can generate trusted results for doctors and patients before scaling to other roles and use cases:

Trust is an important component. Now that ChatGPT is widely available and we are all playing with it, we are seeing how it is not perfect. We don’t need any hallucinations, speculation, or storytelling. We are starting with the basic transcription of what happened in the exam room, and then we can build other functionalities on top of that.

My take

Generative AI hype typically focuses on the ability to generate text, code, or graphics or to power better chatbots. This use case demonstrates the tremendous potential for generative AI to act as a glue between different systems and roles within the enterprise.

The first thing to note is how generative AI can improve the user experience. The record and documentation of what happened are captured in line with the doctor-patient interaction rather than as a secondary process after the fact.

Also, generative AI is not just transcribing what doctors say but tuning it for different audiences. Other doctors are concerned with how this patient compares to others, the range of treatment options and how these turned out. The patient is more concerned with what precautions they should take and the best steps to heal more quickly or at least minimize a disease progression.

Down the road this could reduce to the work to format records for other roles as well. An insurance company will want to know how much a condition cost to treat and how it compares to other healthcare providers. The legal team may want to ensure that all privacy and compliance requirements are documented. This speaks to the broader role that generative AI might play in helping to improve understanding and coordination across product, technology, procurement, and management teams.

It is also essential to provide a trail of breadcrumbs back to the source for each element in a summary. Was a mistake caused by poor transcription or summarization or because someone said the wrong thing? Teams need to give participants a chance to review these summaries to ensure accuracy and minimize the impact of AI and human error on others.

A grey colored placeholder image