Machine Predicts Inpatient Sepsis 5 Hours Sooner

Marcia Frellick

April 10, 2018

ORLANDO — Deep machine learning can help predict sepsis in hospitalized patients an average of 5 hours before they meet the clinical definition, new data show.

"Every hour delay in antibiotic therapy is associated with a 7% to 8% increase in mortality for those with septic shock," said investigator Cara O'Brien, MD, from Duke University in Durham, North Carolina.

"Current tools to identify sepsis don't work very well," she told Medscape Medical News. They only look at the last set of variables, such as the last respiratory rate or last lab result.

Deep learning looks at the trajectory of a patient's data throughout his or her hospital stay.

"It's not looking at a single time point," O'Brien explained. "It's incorporating where the patient was 2 days ago, 1 day ago, 12 hours ago into the risk prediction."

Other tools used to predict sepsis are often compromised by alarm fatigue, said lead investigator Anthony Lin, a third-year medical student at the Duke Institute for Health Innovation.

In fact, Lin and his colleagues found that alarms for sepsis in the Duke system were being canceled 63% of the time.

Current tools to identify sepsis don't work very well.

Sepsis is difficult to detect because the same symptoms can indicate many diseases. Treatment is not difficult; the challenge is finding which patients are septic, Lin explained here at the Society of Hospital Medicine 2018 Annual Meeting.

In their retrospective study, the investigators used a deep-learning tool — which has been used in previous forms of speech-recognition programs, such as Google Translate — to look at fairly common predictor variables, such as vital signs, lab results, medication administrations, and orders.

They looked at all 43,046 adult inpatient admissions at Duke University Hospital from October 1, 2014 to December 31, 2015. An analysis of the millions of data points gleaned from the electronic health records of these patients yielded 83 predictor variables.

The deep-learning tool outperformed the other tools currently available for the early detection of sepsis, Lin reported.

And because the tool uses predictor variables commonly found in electronic health records, it could be used to identify patients experiencing cardiac or respiratory arrest in the hospital or a postoperative infection, the investigators note.

"We're currently investigating how this would play out in cardiogenic shock," Lin said.

The team is planning to launch the deep-learning intervention at Duke later this year. "It's one thing to build a model, but another to implement it in a health system and identify who would answer alarms and who would work up the patient," he noted.

Evolution of Informatics in Medicine

"One of the things that excites me about this research is that it represents an evolution in the way we're using informatics in medicine," said Ethan Cumbler, MD, from the University of Colorado School of Medicine in Denver, who is director of the research, innovation, and vignettes section for the meeting.

"If we can start using this form of artificial intelligence to not just store information within the electronic health record, but to derive from electronic health records new ways of understanding the data, we've created an entirely new way for medical informatics to support clinicians," Cumbler told Medscape Medical News. "That is exciting."

O'Brien, Lin, and Cumbler have disclosed no relevant financial relationships.

Society of Hospital Medicine (HM) 2018 Annual Meeting: Abstract 413603. Presented April 10, 2018.

Follow Medscape on Twitter @Medscape and Marcia Frellick @mfrellick


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.