Scientists Use AI to Improve Patient Care

Electronic health records were widely adopted in the hopes that they would improve patient care and save time. However, doctors often spend more time navigating these systems and less time communicating with patients due to their fragmented interfaces.

Researchers from MIT and Beth Israel Deaconess Medical Center combine machine learning with human-computer interaction to create a better electronic healthcare record (EHR). MedKnowts is a system that unifies both the search for medical records and the documentation of patient information. It's called an interactive interface.

This smart EHR is driven by artificial intelligence and displays personalized, patient-specific medical records whenever a clinician requests them. MedKnowts provides automatic completion for clinical terms, and auto-populates fields containing patient information to aid doctors in their work.

"EHRs were born out of a great enthusiasm. It was clear that it would be beneficial to organize all the information in order to keep track of billing records, provide statistics to the government and make scientific research easier. Few people stopped to think about whether EHRs would be useful for clinicians. Many clinicians feel that they have been pushed to the limit by EHRs for the benefit bureaucracies, scientists, and accountants. David Karger (professor of computer science at the Computer Science and Artificial Intelligence Laboratory CSAIL), is the senior author of this paper.

This research was coauthored by CSAIL graduate student Luke Murray, who is also the lead author, Divya and Monica Gopinath. Steven Horng is an emergency medicine physician and clinical leader for machine learning at Beth Israel Deaconess Medical Center. David Sontag, associate professor of computer science and electrical engineering at MIT, is a member of CSAIL and Institute for Medical Engineering and Science and a principal investigator at Abdul Latif Jamel Clinic for Medical Machine Learning in Healthcare. The Association for Computing Machinery Symposium will present the paper next month.

Problem-oriented tools

Researchers had to think like doctors in order to design an EHR Software that would be beneficial for doctors.

The side panel displays pertinent information from the patient’s medical history and allows for note-taking. This historical information is displayed in cards that focus on specific problems or concepts.

MedKnowts will display a "diabetes cards" when a clinician typed the term "diabetes" into the text. This card contains medications, lab results, and excerpts from past records relevant to diabetes treatment.

Murray states that most EHRs save historical information on separate pages. They list lab values or medications alphabetically, or chronologically. This forces the clinician to search through the data to find what they are looking for, Murray says. MedKnowts displays only information that is relevant to the concept being written about by the clinician.

This is closer to how doctors think about information. This is often subconsciously done by doctors. They will only look at the medication pages that pertain to their current conditions. Murray states that they are helping the doctor automate this process and hopefully free up some time for the more complex part of the process, which is to determine the cause and come up with a treatment plan.

Chips are pieces of interactive text that act as links to other cards. When a doctor writes a note, an autocomplete system detects medical terms such as medication, lab values, and conditions and converts them into chips. Each chip displays a word or phrase highlighted in a specific color according to its category (red for a condition, green for medication, yellow for procedure, etc.).

Autocomplete allows for the collection of structured data about the patient's condition, symptoms and medication use without the physician having to do any additional work.

Sontag said that he hopes the breakthrough will "change how large-scale data can be created for health purposes to study disease progression and assess the effectiveness of treatment."

The researchers developed MedKnowts over a year. They tested it by using the software in an emergency room at Beth Israel Deaconess Medical Center, Boston. They collaborated with an emergency doctor and four hospital scribes, who entered notes into the electronic medical record.

Agrawal said that the deployment of the software in an emergency room, where doctors work in high-stress environments, was a delicate balance act.

"One of our biggest challenges was getting people to change what they do. It is possible to form muscle memory by repeating the same clicks and system over and over again. There is always the question of whether it is worth making a change. We found that certain features were more used than others," she said.

The deployment was also complicated by the Covid-19 pandemic. Researchers had been visiting the emergency room to gain an understanding of the workflow but were forced to stop because Covid-19 meant they were not able to stay in the hospital while it was being deployed.

Over the course of the deployment, MedKnowts proved to be a popular choice among the scribes despite the initial difficulties. The system received an average rating for usability of 83.75 out of 100.

Survey results show that Scribes found autocomplete to be particularly useful in speeding up their work. The color-coded chips also helped them scan their notes quickly for relevant information.

These initial results look promising but the researchers will continue to monitor feedback and work on future MedKnowts iterations.

"What we want to do is make it easier for doctors to get ahead and allow them to accelerate. There are some risks. The purpose of bureaucracy, which is slowing down the process and making sure that all the t's and i's are crossed, is part of its purpose. If a computer is used to cross the t's and tick the i's for doctors, it may be counterproductive to the goal of the bureaucracy which is to make doctors think twice before planning. Karger states that doctors and patients must be protected from the negative consequences of doctors becoming more efficient.

Vision for the long-term

Agrawal states that the researchers are working to improve MedKnowts' machine learning algorithms so that the system can highlight the most important parts of the medical billing software.

They want to consider the different needs of medical users. MedKnowts was designed with the emergency department in mind. This is where doctors typically see patients for their first time. A primary care physician who is more familiar with their patients would probably have different needs.

The researchers are aiming to create an adaptive system where clinicians can participate in the long-term. One example is when a doctor discovers that a particular cardiology term is not in MedKnowts. He or she may add the information to a card which will update the system for all users.

As a way to further deploy the team, the team is looking into commercialization.

"We want tools that doctors can create their own tools. Karger says that while we don't expect doctors learn programming, they may be able to modify any medical software they use to suit their needs and preferences with the right support.

The MIT Abdul Latif Jamel Clinic for Machine Learning in Health funded this research.