At Baptist Health South Florida, which includes 11 hospitals in Miami-Dade, Broward and Palm Beach counties, providers grappled with an overwhelming challenge of managing large amounts of information resulting from the patient-provider dialogue.
THE PROBLEM
A major pain point was the amount of time cardiologists had to invest in documenting patient visits. This manual documentation process was not only time-consuming, but also exacerbated provider burnout and fatigue. As some physicians noted, the longer time spent on clinical documentation meant physicians had fewer opportunities to treat new patients.
“While commercial technologies were available on the market, they came with a hefty price tag,” said Douglas Davila-Pestana, technical AI manager at Baptist Health – another example of a recent trend in AI-specific job titles in healthcare organizations.
“Given the challenging financial environment that many healthcare organizations, including ours, found themselves in, there was an urgent need for a more cost-effective and efficient system to streamline clinical documentation.
“We worked to integrate generative AI into an AI-enabled documentation app that seamlessly combines medical transcription technology with advanced AI, especially large language models,” he continued. “This unique combination allowed the AI ??to quickly generate clinical notes from transcribed patient conversations.”
“Before investing, organizations should weigh the costs of available vendors against the feasibility of developing an in-house solution, taking into account the unique needs and expertise available within their institutions.”
Douglas Davila-Pestana, Baptist Health South Florida
The advantage of the system was its immediacy: physicians could access comprehensive clinical notes shortly after the end of a patient visit, reducing the delay traditionally associated with manual documentation processes, he added.
PROPOSAL
“To combat this manual documentation dilemma, a groundbreaking proposal was made: the use of generative AI, leveraging services such as AWS HealthScribe for medical audio transcription and other tools such as Azure OpenAI, Snowflake and DataRobot,” said Jaymin Patel, data platform and engineering manager at Baptist Health South Florida.
“The idea was to record patient-physician interactions with patient consent, transcribe these recordings into text, and then use a large language model to generate clinical summaries in a clinical SOAP format,” he explained.
“It was expected that this automation process would dramatically reduce documentation time to just two to five minutes after the visit. Additionally, through the integration with Snowflake, which supports our data lake, data warehouse and reporting layer, it was proposed that patient and appointment data can be easily pulled from our data warehouse platform to further enhance and personalize these summaries.”
GPT-4, supported by research published in the Journal of the American Medical Association, was highlighted as a promising LLM, given its demonstrated skill in suggesting medical diagnoses. All of these measures were aimed at improving the patient experience, improving clinical productivity, improving operational efficiency and leading to significant cost savings.
“Additionally,” Patel said, “the intention was for the data generated by the application to be stored in the Snowflake data lake layer for further analysis and application improvements.”
MEETING THE CHALLENGE
The technology journey started with recording the interactions between patient and doctor. These recordings were sent to the AWS HealthScribe service for transcription. Once transcribed, the text was entered into the large language model, which includes GPT-4. By interpreting the transcript, this model generated concise summaries in clinical SOAP format, which were then verified by physicians for accuracy.
“Cardiology physicians, along with their beta testing group of physicians, were the main users,” Davila-Pestana noted. “For the tech stack, AWS Lambda and AWS S3 were critical in building the AI ??service. The moment an audio transcription file was generated, the AI ??Lambda Service built on AWS was activated.
“It was expected that this automation process would dramatically reduce documentation time to just two to five minutes after the visit.”
Jaymin Patel, Baptist Health South Florida
“Integration with Snowflake was essential to extract patient appointment data,” he continued. “Additionally, once the summary was edited and approved by the physician, it would be exported and integrated with our electronic health record system. This entire process ensures timely, accurate, and standardized clinical documentation.”
RESULTS
This use of AI yielded significant results in three key areas, Davila-Pestana explains:
-
Time saving. The group of physicians notes a reduction of several minutes per patient interaction compared to the total time spent on clinical documentation. This new efficiency gives them more bandwidth to treat new patients, improving the patient care experience.
-
Improved clinical productivity. With the AI-assisted summaries, doctors no longer have to wait for hours as with previous tools. The summaries can now be generated in just minutes, resulting in faster turnaround times and more streamlined operations.
-
Cost efficiency. By taking this approach, Baptist Health South Florida avoided the significant costs of commercially available options, leveraged the organization’s intellectual resources and technology, and capitalized on the efficiencies of AI.
ADVICE FOR OTHERS
“For healthcare providers considering the integration of generative AI technology, it is critical to view AI as a tool that augments rather than replaces human input in clinical documentation,” Davila-Pestana advises. ‘AI can deliver enormous efficiency gains, but blind trust can lead to oversight.
“Ensuring a human-in-the-loop verification process, where physicians review and validate the accuracy of AI-generated summaries, is critical to maintaining the integrity of patient records,” he added. “Additionally, before investing, organizations should weigh the costs of available vendors against the feasibility of developing an in-house solution, taking into account the unique needs and expertise available within their institutions.”
Finally: any AI implementation must prioritize patient consent and data security to maintain trust and comply with healthcare regulations, he said.
Follow Bill’s HIT coverage on LinkedIn: Bill Siwicki
Email him: [email protected]
Healthcare IT News is a HIMSS Media publication.