New Tool for Predicting C. Difficile Infections Shows Promising Results

Shenoy-headshot.jpg

Erica Shenoy, MD, PhD

For patients in hospital and healthcare settings, a Clostridium difficile (C. difficile) infection is a complication that can result in serious complications and even death.

C. difficile is caused by a bacterium and the symptoms of infection include diarrhea, fever and severe abdominal cramps. While some cases may be mild, some can be fatal—particularly in elderly, very sick or other vulnerable patient population.

Antibiotic use can increase the risk of C diff infections because antibiotics can kill off beneficial bacteria that normally keep C. difficile at bay.

C diff spreads through contact with patients or contaminated surfaces, which makes early diagnosis important—both so patients can get the right treatment early in the course of disease to reduce the severity but also so patients can be isolated to prevent the spread to others.

Approximately 300,000 patients are diagnosed with C. difficile in health care settings each year and it causes 30,000 deaths. The infections are expensive as well, costing acute care facilities close to $5 billion to treat and contain each year.

Despite substantial efforts to improve diagnosis and treatment of C. difficile, infection rates have continued to increase, says Erica Shenoy, MD, PhD, an infectious disease physician and Associate Chief of the Infection Control Unit at Mass General. “We need better tools to identify the highest risk patients so we can target both prevention and treatment outcomes to reduce further transmission and improve patient outcomes.”

To address this challenge, Shenoy collaborated with researchers from Massachusetts General Hospital, the Massachusetts Institute of Technology, and the University of Michigan Hospital system (UM) to create a machine learning-based tool that can predict the risk of C. difficile infection for patients during their admission.

Machine learning is the process of using computer algorithms to detect patterns in data. The algorithms used to calculate infection risk were built by analyzing large quantities of de-identified patient data (i.e. demographics and medical history), details of admission and hospitalization, and the likelihood of exposure to C. difficile from 257,000 patients admitted to Mass General and the UM hospital system.

The team then created two predictive models—one customized to each institution. While similar in approach, the models created were unique to each institution, allowing for institution-specific risk to be calculated.

Shenoy explains that the ability to customize the tool was critical to the project, as infection rates are affected by unique institutional characteristics such as patient populations, hospital layout, testing and treatment protocols, antibiotic use, and how health care providers interact with the electronic health record.

The team’s new institution-specific models proved to be successful in predicting which patients would ultimately be diagnosed with C. difficile. In half of the cases predicted correctly, the model identified the patients who went on to develop C diff at least five days before diagnostic samples were collected.

The study was published last month in Infection Control and Hospital Epidemiology, and initially presented last October at ID Week 2017. Last month, Dr. Shenoy presented the team’s research as part of the First Look Series at the Partners World Medical Innovation Forum.

The research team has also made the algorithm code freely available here for other investigators to review and adapt for their own institutions. The team hopes that other researchers will use the code to explore its use in their facility.

“We strongly recommend that interested institutions assemble the right team before embarking on this sort of project—the expertise of computer scientists, informatics specialists, infectious disease and infection control, are all pre-requisites for success,” says Dr. Shenoy.

Leave a Reply

%d bloggers like this: