Reposted with permission from Pulse+IT.
Global electronic medical record vendor MEDITECH has collaborated with Google Cloud to develop an AI-powered search and summarisation capability in MEDITECH’s Expanse EMR to enable care teams to find the information they need using Google’s familiar search function.
Powered by Google Cloud’s foundational and generative AI capabilities, the Expanse search and summarisation tool uses natural language processing and a large language model to delve into the medical record and extract data that is presented in context to the user from within their workflow.
The new tool has been trialed at Mile Bluff Medical Center in Wisconsin in the US on a cloud version of the Expanse EMR, hosted in the Google Cloud, and is shortly to go into trial for the on premises version of Expanse before a wider release to worldwide MEDITECH users in the next year.
Google’s senior clinical specialist Peter Clardy said the search and summarisation capabilities use natural language processing to intelligently pull information from structured and unstructured data, including scanned documents, faxes, legacy data and even handwritten notes.
Instead of relying on specific word matches, the search capabilities use terminology, abbreviations and medical language that make sense to clinicians, and has the ability to understand the context and concepts behind the user’s search.
Dr Clardy, a pulmonary and critical care physician by training, said the idea for the search and summarisation tool began some time ago as part of Google’s foundational research program, “some of which becomes a product in the end, and much of which either becomes open source or published for academic reasons”.
Dr Clardy said Google Health began working in the medical records space to focus on a few challenges.
“One is the challenge around data fragmentation and the fact that data lives in lots of different areas, and patients have these digital crumbs that are sort of scattered about,” he said.
“Another is that although the language of medicine and healthcare is very specific, and there’s lots of health data that is structured, a lot of medical information is locked up in what we would call unstructured, natural language text, the things that show up in medical notes.
“So there’s the data fragmentation problem with the unstructured data, and then data overload, where providers are overloaded with data.
“And the third piece: this is the layer that we’re really excited about, and that’s where it’s not just AI, but generative AI. And that’s the ability to summarise results, either of a search or of a record organised one way or another. That’s the origin of the idea and the core elements.”
Google Health began working with MEDITECH to take this foundational technology and develop it and scale it up for their MEDITECH customers, he said. “A lot of the work, a lot of things that are central to the solutions that we are privileged to develop with MEDITECH, has to do with making information available.”
Pilot expands beyond physicians
The tools were piloted in January at Mile Bluff Medical Center, which has 40 inpatient beds and around 17 different departments, starting with 130 users. About 170 are now using the tool for a range of different uses.
MEDITECH senior director of field marketing operations, Christine Silva, said it was not just physicians at the hospital who could see a use for it, but also the health information management (HIM) team and the infection control team.
HIM was not necessarily in the plans for an initial rollout, but asked to be part of it, Ms Silva said.
“For the HIM department specifically, they are using this tool on about 50 patients a week, and they’re saving 17 hours a week as a team, so there’s significant time savings,” she said.
For doctors, the search function has unlocked tasks such as preparing for new patient visits or seeing patients after they had been hospitalised, or prepping for visits when have had a busy session with patients over the course of the morning.
“They are using the search of the medical record in the process of care, rather than doing their normal browsing, which is the typical way of looking for information in the medical record,” she said.
“In addition to the search capabilities, another element of the solution is a longitudinal, intelligent view of patient conditions. That is useful if you’re seeing a new patient, or seeing a patient in a consultation or after a hospitalisation, or seeing a patient after a long interval, or trying to prep a number of patients all at once.”
Clinical champion at Mile Bluff, primary care physician assistant and team lead for the project Randy Brandt, uses the search functionality and summarisation tool as an educational resource when he is sitting with patients. Google presents the data in the Expanse Chart Viewer and has great graphing capability, Ms Silva said.
“Often, Randy will turn the screen to his patients and say, ‘Can you see that when you’re taking your medication as prescribed and you’re monitoring your diet, or you’re following the steps that the physician has prescribed to you, you can see how your direct data trends in the right direction.”
Dr Clardy said one of the other areas that had been interesting was the ability to search in context.
“For example, I would use a very specific search term such as TSH, which means thyroid stimulating hormone, and that’s a very specific lab result. But for a provider to see that result in isolation, it’s kind of tricky to interpret.
“If you do that sort of search with the tools that we’ve created, you’ll see the context that surrounds TSH. It will be the result itself, the result over time, thyroid medications and mentions of the lab result and the associated meds in notes. With one search term, you get this broader view.
“It handles something that we’ve often referred to as medical FOMO [fear of missing out], meaning clinicians browse data because they don’t want to miss anything that’s contextually important for the specific case.
“It was really useful to see both Mile Bluff and how MEDITECH thinks about organising information that we’re able to give a sort of separate layer to search results and the summarisation, and that’s presented to the clinician in context and within their normal workflow.”
Trust and grounding
One major issue AI faces in healthcare and in other applications is the issue of trust. Dr Clardy said that as an ICU doctor, if he was seeing a patient and they had a long record, if the summary wasn’t grounded back into the medical record, it wouldn’t save any time.
“All of those things that I need to to validate the paragraphs will involve me looking at exactly the things I would have needed to essentially generate that paragraph in my own head,” he said.
“When we develop these tools, this concept of grounding is really central. It’s almost as if you’re kind of showing in reverse the way the AI tools are generating the summary. So for any sentence, you can go back to the specific sections of a note that are used to create that sentence.
“Each practitioner will have their own learning curve and trust curve, but we think that showing where the information comes from is really important.”
Both the search and summarisation are accessed right from the toolbar in the Expanse desktop, and the next iteration of the tool will involve not just patient conditions, but the ability to click on the condition and see the additional information within the patient notes.
Dr Clardy emphasised that Google has not used MEDITECH data or live patient records to train the large language model.
“The data belongs to patients and their providers, and MEDITECH has a very special relationship with those patients and providers in terms of handling of that data.
“Google takes data privacy and security very seriously, and we’ve used different data types over a long period of time to develop some of our healthcare specific models.
“We don’t use any Mile Bluff data or MEDITECH data to build any of the foundational models. The data that we use is generally either publicly available data or data that is de-identified by a third party before it will be used in model training.
“Now the opportunities going forward for users at a given organisation is to develop models that are trained on their own data, and to continue and to develop and refine that, I think, is a longer vision, but in terms of data coming in for training, that’s not something that we do.”
Results
Ms Silva said the most recent results from the trial show that the clinicians using the tool are saving anywhere from five to 10 minutes per patient, so an average time saving per physician of about seven and a half minutes per patient.
“If I’m a physician and I’m seeing 10 patients in the morning, I have a little over an hour back in my day,” she said.
As well as doctors and health information managers, the infection control team, which has to wade through copious amounts of data for reporting, is also saving about 40 hours per month.
“The infection control team, looking at surgical site infections, have a very clear but very constrained set of terms that they need to find in natural language across messy records from outside and from other sources, as well as a lot of records from university hospitals to which they refer patients,” Dr Clardy said. “They’re using the tool to condense things down.”
Another use case is the ability to take on extra and late consults. One physician using the tool is a sleep specialist who has little calendar space.
Dr Clardy said the major rate limiting step for that doctor seeing new patients is the exhaustive search that he needs to do to find sleep studies and other outside information. He can now quickly find the pieces of information he needs, and this has allowed him to see new patients, having lowered the barrier of entry for patients.
Ms Silva said 86 per cent of users reported they were either somewhat satisfied or very satisfied with the tool, and 76 per cent of those users said it was very or extremely helpful in performing tasks that they need to do on a daily basis. And 91 percent of users feel it is saving them time.
Ms Silva also said the tool was incredibly easy for users to learn how to use. “All we needed to do was spend just a few minutes training them on how to use the tool. Because when you think of Google, you immediately think of search capabilities, and if you know how to use Google, you can easily use the tool.
“It’s really simple to onboard clinicians. You can sit with someone for a couple of minutes, show them how to use it, and then they’re off and running.”
MEDITECH is expecting to onboard two new customers shortly who use an on prem version of MEDITECH. Mile Bluff is a MEDITECH as a service (MaaS) customer, hosted in the Google Cloud.
She said she hoped to have those customers live by the end of 2024 or early 2025 and then from then on, in both the MaaS environment and the on prem environment, it will start to be rolled out worldwide.