MADISON, Wis. – The volume of electronic medical notes per patient available to emergency medicine providers has grown 30-fold in 17 years, according to a new study from the University of Wisconsin School of Medicine and Public Health.
This presents an emergency care team with a novel’s worth of words they must sift through to understand the medical history of patients they are often meeting for the first time.
When a patient is seen in the emergency department, physicians review available records in a process often referred to as a “chart biopsy.” Physicians skim through previously recorded notes to make sure they are aware of medical conditions that may pertain to a patient’s visit.
As records store more and more data, this task has become increasingly difficult, according to Dr. Brian Patterson, associate professor of emergency medicine, UW School of Medicine and Public Health, emergency medicine physician, UW Health and lead author of the study.
The study tells a dramatic story of providers adrift on an ocean of notes within electronic health records, he said.
“That’s why we called the paper, ‘Call me Dr. Ishmael,’ after Herman Melville’s classic seafaring novel, ‘Moby Dick,’” Patterson said.
Patterson and his collaborators analyzed data from about 731,000 patient visits between 2006, when UW Health began to use electronic health records, and January 2023. The visits took place at two UW Health emergency rooms; University Hospital, which is the larger of the two facilities and has been in operation for many decades, and East Madison Hospital, which opened in 2015. In 2006, the median patient had five notes; by 2022, the median patient had 359 notes.
By the end of the study period, one in five patients arrives in the emergency room with a chart the size of Moby Dick, which comprises more than 206,000 words, according to the study. Nearly 4% of patients have charts the size of Leo Tolstoy’s “War and Peace,” a saga that clocks in at more than 560,000 words.
“Prior to 2010, reading a chart biopsy for the average patient involved skimming several notes, not easy but possible within a few minutes,” Patterson said. “This task has become more burdensome, to the point where it is often impossible to get a handle on a patient’s history within the time constraints of an emergency visit.”
These notes are valuable and important to keep for a patient’s comprehensive medical history and care. Prior to these notes, emergency medicine providers often didn’t have the valuable information they needed to care for patients. Within the context of modern emergency rooms, it’s about being able to comprehend the key information efficiently, he said.
“Unfortunately, we’ve become victims of our own success, and in many cases, we’ve shifted from not knowing enough about a patient to having an overwhelming amount of information about a patient to review in one sitting,” Patterson said.
While emergency providers must synthesize the information quickly to decide how to treat the patients, the chart biopsy process is not necessarily easier for physicians who treat patients that end up being admitted to the hospital after an emergency room visit, according to Patterson, because admitted patients tend to have larger medical charts.
The research team used the programming language Python to help rate the size of the charts, using the total number of notes, word count and tokens, which are clusters of characters used by artificial intelligence models to represent fundamental units of text.
While information in electronic health record notes improves patient care, researchers explained that clinicians need better ways to synthesize the data because they currently risk missing critical pieces of data. Artificial intelligence large language models that can analyze charts may help providers navigate this expanding body of data, according to Frank Liao, senior director of digital health and emerging technologies at UW Health, and a co-author of the study.
“These large language models can be trained to generate concise and relevant summaries of patient data, allowing physicians to quickly grasp essential information without sifting through extensive notes, potentially reducing the cognitive load on the physician,” he said.
One limitation is the current constraints on the number of tokens that can be analyzed. For example, when this study was published, GPT-4 Turbo had an upper limit of 128,000 tokens while the median size of electronic health record notes for a patient aged 65 or older contains almost 139,000 tokens.
The study was published in the Journal of the American Medical Informatics Association.
Other members of the UW research team include Daniel Hekman, Dr. Azita Hamedani, Dr. Manish Shah, and Dr. Majid Afshar.
