Decision-making forms the core of hospital patient care, involving an array of clinicians whose duties span diagnosis, treatment and resource allocation. The complexity of these interrelated decisions makes it challenging for physicians, nurses and other caretakers to connect all the dots in real time.
Shengpu Tang, assistant professor of computer science at Emory University, is developing AI tools to identify, validate and transmit key data needed to most effectively support health care workers in decision-making processes.
“The end goal is to improve patient care and patient outcomes,” Tang says.
JAMA Network Open published the results of Tang’s latest collaborative project: the first AI guidance deployed in a hospital to support best practices for preventing the spread of dangerous infections of Clostridioides difficile.
A medical illustration of Clostridioides difficile bacteria by the Centers for Disease Control and Prevention.
Tang is first author of the JAMA Network Open paper, based on research done at the University of Michigan and its affiliated academic medical center, Michigan Medicine.
Analysis by the researchers found that the new AI-guided protocol significantly reduced antibiotic prescriptions at Michigan Medicine — a factor that increases infection risk for vulnerable patients — with 10-15% fewer days on antimicrobials. Importantly, the reduction of days on antimicrobials did not increase length of stay, readmission rate or mortality among patients. The already low incidence of Clostridioides difficle trended downward during the study, but that reduction did not reach statistical significance.
Tang did the work as a PhD student in the lab of Jenna Wiens, senior author of the paper and associate professor of computer science and engineering at the University of Michigan. Krishna Rao, associate professor of internal medicine at the University of Michigan Medical School, is co-senior author of the paper.
An urgent threat
The bacterium Clostridioides difficle, or C. diff, is an urgent antibiotic resistance threat, causing half a million illnesses in the United States each year and nearly 30,000 deaths, according to the Centers for Disease Control and Prevention (CDC).
“C. diff is one of the most common health care-associated infections seen in hospitals,” Tang says. “It’s a very stubborn problem.”
When outside the body, the bacterium forms spores that can remain on surfaces for months and are resistant to many cleaning products, including alcohol-based hand sanitizer. These factors make C. diff particularly dangerous in hospital settings. In people whose immune systems are already compromised due to illness, a C. diff infection can cause severe diarrhea and gut inflammation.
Specifically, hospital patients taking antibiotics are 10 times more likely to get a C. diff infection. The medications wipe out the indigenous microbes in the gut, which naturally form a barrier against such invaders.
The current paper is part of a project more than a decade in the making. In the early stages, the research team built a predictive model to pinpoint which patients are at the greatest risk of C. diff infection using past hospital records.
The machine learning model trained on factors including medications, lab results, previous hospitalizations, comorbidities, demographics and even proximity to other infected patients in the hospital. Applied to a new set of never-before-seen patients, the predictions aligned with true patient risk, showing that the model worked.
Determining the optimal risk threshold
Tang began working on the project in late 2021, after the model was built, and helped lead its deployment. Initially, he focused on coding and debugging software and ensured upstream data sources connected properly into the electronic record database for patients.
Later, Tang conducted simulation studies to determine the optimal risk threshold for sending a high-risk alert, based on conversations with other researchers on the team. He also coordinated with hospital IT personnel to ensure the smooth implementation of the model into hospital workflows. After the study concluded, he led the statistical analysis of the results.
The AI model generated a daily risk score for patients. Alerts are sent for patients whose risk exceeds the selected threshold. The alert appears on the high-risk patient’s electronic health record, visible to physicians, nurses and pharmacists. The AI guidance requires posting a sign on a high-risk patient’s room, asking all visitors to wash their hands with soap and water before entering.
Guidance also includes reducing the use of high-risk antibiotics and testing for penicillin allergies. Many patients labeled as allergic to penicillin lose the allergy over time, opening a new class of antibiotics that lowers the patient’s risk of getting infected with C. diff.
The acute awareness across the hospital of high-risk patients led to measurable changes in antibiotic regimens, which reduce the risk of getting infected.
A team of intensive care nurses even implemented an unexpected use for the patient risk score in their own workflow. When assigning rooms, the charge nurse ensured a nurse caring for a patient with an active infection was not also assigned to a high-risk patient.
“This project reflects the collective effort of a team of clinicians, pharmacists, engineers and scientists working together to translate AI into real-world impact,” Tang says. “At Emory, I look forward to continuing this line of work and exploring innovative ways AI can help improve patient care.”
Patrica DeLacey contributed to this report.