The healthcare industry has embraced electronic health records (EHRs) with expectations of enhanced efficiency and improved patient outcomes. However, cumbersome interfaces and time-consuming data entry have forced medical professionals to spend excessive time navigating complex systems rather than focusing on patient care.
In a groundbreaking collaboration, researchers from MIT and Beth Israel Deaconess Medical Center are harnessing the power of artificial intelligence to revolutionize EHR systems. Their innovative creation, MedKnowts, represents a significant leap forward in intelligent patient data management, seamlessly integrating medical record retrieval and documentation within a single, intuitive interface.
This advanced smart EHR system leverages cutting-edge machine learning algorithms to dynamically present customized, patient-specific medical information precisely when clinicians need it. MedKnowts dramatically enhances workflow efficiency through intelligent autocomplete functionality for clinical terminology and automated population of patient data fields.
"When EHRs were first introduced, there was tremendous optimism about organizing health information for billing, government reporting, and research purposes. Unfortunately, few considered whether these systems would actually serve clinicians' needs. Many healthcare providers feel burdened by EHRs designed primarily for administrative and research benefits. Our project fundamentally reimagines how EHRs can directly benefit clinicians," explains David Karger, computer science professor at CSAIL and senior author of the research paper.
The research team includes lead author Luke Murray, along with fellow CSAIL graduate students Divya Gopinath and Monica Agrawal. Additional contributors include Steven Horng, emergency medicine physician and clinical machine learning lead at Beth Israel's Center for Healthcare Delivery Science, and David Sontag, MIT associate professor and principal investigator at the Abdul Latif Jameel Clinic for Machine Learning in Health. The findings will be presented at the upcoming ACM Symposium on User Interface Software and Technology.
Problem-Focused Design Approach
Creating an EHR that genuinely serves physicians required the research team to adopt a clinician's perspective.
They developed an innovative note-taking editor featuring a side panel that dynamically displays relevant patient history information. This historical data appears as contextual cards focused on specific medical problems or concepts.
For example, when MedKnowts detects clinical terms like "diabetes" as a clinician types, the system instantly generates a specialized "diabetes card" containing relevant medications, laboratory results, and historical notes pertaining to diabetes management.
Traditional EHR systems typically organize historical information across separate pages, listing medications or lab results alphabetically or chronologically. This approach forces clinicians to manually search through extensive data to find relevant information. In contrast, MedKnowts selectively displays only information pertinent to the specific clinical context being documented.
"This design mirrors how physicians naturally process information. Doctors often subconsciously filter through medication lists, focusing only on those relevant to current conditions. We're automating this filtering process, potentially reducing cognitive load so clinicians can dedicate more mental energy to complex diagnostic reasoning and treatment planning," Murray explains.
The system employs interactive text elements called "chips" that function as links to related information cards. As physicians compose notes, the intelligent autocomplete system identifies clinical terms—including medications, laboratory values, or conditions—and converts them into color-coded chips (red for medical conditions, green for medications, yellow for procedures, etc.).
Through this seamless autocomplete functionality, MedKnowts automatically captures structured data regarding patient conditions, symptoms, and medication usage without requiring additional effort from healthcare providers.
Sontag expresses optimism that this advancement will "fundamentally transform how large-scale health datasets are created for studying disease progression and evaluating real-world treatment effectiveness."
Real-World Implementation
Following a year-long iterative design process, the research team deployed MedKnowts in the emergency department at Beth Israel Deaconess Medical Center in Boston. The implementation involved one emergency physician and four hospital scribes responsible for EHR documentation.
Implementing the system in a high-stress emergency environment presented significant challenges, according to Agrawal.
"One of our biggest hurdles was encouraging workflow changes. Physicians using the same system repeatedly develop muscle memory for specific navigation patterns. When introducing changes, we constantly questioned whether the benefits justified the adaptation required. We observed that certain features gained more traction than others," she notes.
The COVID-19 pandemic further complicated deployment efforts. The research team had to discontinue in-person observations of emergency department workflows and couldn't be present during the actual system implementation due to hospital access restrictions.
Despite these challenges, MedKnowts gained popularity among scribes during the one-month deployment period, receiving an average usability rating of 83.75 out of 100.
Survey results indicated that scribes particularly valued the autocomplete functionality for accelerating documentation tasks. Additionally, the color-coded chip system enabled rapid identification of relevant information within clinical notes.
While these initial results are promising, the researchers approach future iterations of MedKnowts with careful consideration of user feedback.
"Our goal is to streamline clinical workflows and enhance physician efficiency. However, this approach carries inherent risks. Administrative processes often serve as deliberate safeguards, ensuring thorough documentation and thoughtful decision-making. When AI systems automate these checks, we must consider how to protect both providers and patients from potential consequences of increased efficiency," Karger cautions.
Future Vision
The research team plans to enhance MedKnowts' machine learning algorithms to improve the system's ability to highlight the most relevant portions of medical records, according to Agrawal.
They also aim to address the diverse needs of different medical specialties. While MedKnowts was initially designed for emergency department settings—where clinicians typically encounter patients for the first time—primary care physicians with established patient relationships likely have different requirements.
Looking further ahead, the researchers envision developing an adaptive system that allows clinicians to contribute improvements. For instance, a physician might identify missing cardiology terminology and add this information to a relevant card, automatically updating the system for all users.
The team is exploring commercialization opportunities to expand deployment of the technology.
"We aim to create tools that empower physicians to customize their own digital environments. While we don't expect doctors to become programmers, with appropriate support, they might significantly tailor medical applications to their specific needs and preferences," Karger concludes.
This research received funding from the MIT Abdul Latif Jameel Clinic for Machine Learning in Health.
