How Jenna Wiens and her team at the University of Michigan’s AI Laboratory are tackling hospital inpatient infections with the help of big data and AI


An increasing number of hospital inpatients end up suffering from infections that they didn’t have at the point of admission. Clostridium difficile is a diarrhea and colitis causing bacterium that is easily transmitted in healthcare facilities and the source of nearly half a million inpatient infections in the US every year. According to a Centers for Disease Control and Prevention (CDC) report released in 2015, approximately 29,000 patients would die 30 days from initial diagnosis.

A conventional way to prevent C. difficile involves examining a small number of risk factors and building a model based on them. Jenna Wiens, Assistant Professor of Computer science and Engineering at the University of Michigan and Lead of the Machine Learning for Data-Driven Decisions (MLD3) research group at the Michigan AI lab believes the approach can be personalized and made more efficient with the help of big data and AI.

“I think there’s a bigger cost of not using the data,” Wiens says. “Patients are dying when they seek medical care and acquired one of these infections. If we can prevent those, the savings are priceless.” Wiens and her research team built a computational model using patients’ information including lab results, prescriptions, and procedures they have undergone to highlight specific risk factors for C. difficile at respective hospitals. This facilitates the identification of the most vulnerable patients and proposed interventions based on risks.

The model is also helpful for researchers conducting clinical trials for the generation of new antibiotics. Conducting trials on C. difficile has proved to be difficult in the past because infections come too fast to involve patients. “I think to get all the value we can out of data we are collecting, it’s necessary to leverage data mining and machine learning,” Wiens adds. “Our approach essentially throws everything in that’s available.”

This new model can identify and intervene at-risk patients at least five days in advance. Some of the interventions can be cleaning the environment or determining who to test for the infection, all of which would improve patient outcomes by decreasing length of stay, overall transmission, and costs. Wiens’ effort finds her on the Forbes 30 under 30 Science and Healthcare list in 2015 and MIT Technology Review’s list of innovators under 35 in 2017.

With an aim to augment clinical care, Wiens and MLD3 also work closely with other research centers to solve healthcare problems and utilize longitudinal data – following patients over a period – for modeling patient trajectories to better understand the progression of various diseases including Type 1 diabetes, Alzheimer’s and cystic fibrosis.

Some of the recent work includes a collaboration with the Michigan Integrated Center for Health Analytics and Medical Prediction (MiCHAMP) to develop and improve on a model that can predict the onset of acute respiratory distress syndrome (ARDS) throughout a hospitalization better than the best clinical techniques currently used, by using routinely collected electronic health record data.

During the ongoing COVID-19 pandemic, Wiens, MLD3 and their colleagues at the Institute for Health Policy Innovation and Michigan Medicine evaluated and built models to predict patient outcomes and resource utilization in patients that tested COVID-19 positive to guide clinical and operational work. They sought to find patients with the greatest risk of complications or adverse outcomes from the disease. Particularly, those who might need respiratory support including a mechanical ventilator, cardiovascular support and those who would likely die in hospital from the disease. Wiens believes the resulting model could help hospitals manage a surge of incoming patients and anticipate needs in advance.

Even though Wiens and MLDS3 have many projects that are in different stages of development, they feel there remains an ultimate challenge: What do we have to do to make AI work for healthcare, clinicians and patients? This gets into questions of implementation science, translation, and making sure medical AI tools are accurate, interpretable, and robust.

“We need to integrate these models into the electronic health records to automatically compute a patient’s risk for a particular outcome,” Wiens explains. “But then the question becomes, who’s the right person to show those data to? Is it the physician, the nurses, the hospital’s infection prevention and control team or the antimicrobial stewardship team? That’s where it gets more application-specific and also where the clinicians can help determine what’s most actionable in the model, how to refine the model, and how to best utilize it.”

“We have seen a tremendous uptick in opportunities to address various issues in healthcare with recent advances in terms of our ability to collect and store health data, clinical data, and by working together in interdisciplinary teams we can begin translating back to the bedside and improving patient outcomes. That’s the fundamental goal.”