In emergency medicine, advances in data science and machine learning may speed detection and treatment of sepsis.
University of Washington
Click for a full size image
Machine-learning tool mines health record for sepsis risk
Emergency medicine big data model better identifies patients needing more aggressive sepsis treatment
A newly developed machine-learning technique examines electronic medical records to predict the danger of severe sepsis in emergency department patients.
Sepsis occurs when a patient is overwhelmed by an infection. It can progress to tissue damage, organ failure or circulatory collapse. More than half of hospital deaths in the United States are related to sepsis.
Sepsis may start with any infection anywhere in the body. Diagnosis can be difficult at first because the initial distress is indistinguishable from other types of shock. The combination of symptoms may include fever, chills, fast heartbeat, rapid breathing, pale skin, confusion, disorientation, and lethargy.
Sepsis needs to be recognized as soon as possible, before it becomes life-threatening or causes permanent harm.
Because studies have shown that early detection and treatment of severe sepsis improves patient outcomes, several groups of researchers have previously designed decision aids to help assess the severity of a patient’s illness. These tools, derived from clinical rules or point-scoring systems, try to identify which patients may die from sepsis who might benefit from more aggressive treatment.
Historically, however, these aids have underperformed in actual practice. They are underused because they require additional data collection or entry by providers who are busy caring for emergency patients. The decision tools are based on tradition statistical models that struggle to capture the richness and complexity of medical data.
Yale University School of Medicine and UW School of Medicine researchers have developed a new method that uses random forest modeling, a machine learning technique, to evaluate patients electronic health records and warns if death from sepsis is looming. The machine learning technique incorporated into their decision tool is similar to the way Netflix predicts its customers’ movie preferences.
The machine-learning approach outperformed several of the best existing clinical decision aids for sepsis severity. A detailed study evaluating the methods appears in the Mardh edition of journal, Academic Emergency Medicine. (DOI: 10.1111/acem.12876)
While traditional approaches concentrate on five or six clinically important variables, machine learning techniques, such as random forest modeling, can harness a large amount of local data available in the electronic medical record. The new method could look at more than 500 variables from over 5,000 emergency department visits.
Compared with traditional modeling, the improved predictive abilities of this new approach could result in properly identifying an additional 200 patients per 5,000 as having severe septic shock, jhe researchers concluded.
Prompt and appropriate treatment is critical in such severely ill emergency patients.They would otherwise be missed through traditional decision aids, the researchers noted.
“Machine-learning techniques can analyze complex interactions among data collected as part of routine care, and hold great promise in the push for building healthcare systems that learn from the data collected.” said Dr. Richard A. Taylor, assistant professor of emergency medicine at Yale University and lead author of the study.
According to senior author Dr. M. Kennedy Hall, acting instructor of emergency medicine at the University of Washington: “It’s all about making the data work for the patient. Our findings demonstrate the potential to develop these approaches to improve predictions across a wide variety of domains within emergency care.”
source : University of Washington