It’s similar to the “precogs” who predict crime in the movie Minority Report, but for sepsis, the deadly response to infection. That’s how Tim Buchman, director of the Emory Critical Care Center, described an emerging effort to detect and ward off sepsis in ICU patients hours before it starts to make their vital signs go haywire.
As landmark clinical studies have documented, every hour of delay in giving someone with sepsis antibiotics increases their risk of mortality. So detecting sepsis as early as possible could save lives. Many hospitals have developed “sniffer” systems that monitor patients for sepsis risk. See our 2016 feature in Emory Medicine for more details.
What Shamim Nemati and his colleagues, including bioinformatics chair Gari Clifford, have been exploring is more sophisticated. A vastly simplified way to summarize it is: if someone has a disorderly heart rate and blood pressure, those changes can be an early indicator of sepsis.* It requires continuous monitoring – not just once an hour. But in the ICU, this can be done. The algorithm uses 65 indicators, such as respiration, temperature, and oxygen levels — not only heart rate and blood pressure. See below.
Example patient graph. Green = SOFA score. Purple = Artificial Intelligence Sepsis Expert (AISE) score. Red = official definition of sepsis. Blue = antibiotics. Black + red = cultures. Around 4 pm on December 20, roughly 8 hr prior to any change in the SOFA score, the AISE score starts to increase. The top contributing factors were slight changes in heart rate, respiration, and temperature, given that the patient had surgery in the past 12hr with a contaminated wound and was on a mechanical ventilator. Close to midnight on December 21, other factors show abnormal changes. Five hours later, the patient met the Sepsis-3 definition of sepsis.
As recently published in the journal Critical Care Medicine, Nemati’s algorithm can predict sepsis onset – with some false alarms – 4, 8 even 12 hours ahead of time. No predictor is going to be perfect, Nemati says. The paper lays out specificity, sensitivity and accuracy under various timelines. They get to an AUROC (area under receiving operating characteristic) performance of 0.83 to 0.85, which this explainer web site rates as good (B), and is better than any other previous sepsis predictor.
“To our knowledge, this is the first study to demonstrate acceptable performance of a sepsis prediction algorithm over incrementally longer time windows,” the authors write.
The algorithm was “trained” on three years of Emory data, including 31,000 ICU admissions, and then validated on a publicly available data set from Beth Israel Deaconess Medical Center. The algorithm is headed for prospective clinical testing, and that’s something Lab Land has been talking with Buchman about. The question is how to handle “alert fatigue”; nurses and physicians working at the bedside get lots of beeps and warnings already. The answer, Buchman thinks, is to create an additional level of care, including a remote team of safety officers whose primary charge is to detect and evaluate evolving threats to each patient.
Here are two other 2017 papers from Nemati and colleagues on using predictive algorithms on sepsis:
About the author
