Dr. David Hagar treats dozens of patients each day at the intensive care unit at John Hopkins Hospital in Maryland. One of his patients was almost perfectly healthy except for having low blood pressure. Within four hours, the patient died of septic shock.
A group of computer scientists at John Hopkins University partnered with Hagar, and created an algorithm that can predict septic shock and give clinicians more time to treat someone at risk.
Septic shock, which is the third level of sepsis, is difficult to predict. Sepsis is a severe immune system response triggered by an infection. If untreated, inflammation spreads throughout the body and can clot vessels. This blocks the blood flow to organs which can cause failure. Physicians treat septic shock by contracting vessels or giving the patient fluids to maintain normal blood pressure and supply blood back to vital organs.
The warning symptoms of an attack can progress over weeks or hours.
The computer system sifted through a dataset of over 16,000 patient electronic health records, which includes a historical profile of blood pressure, heart rate, and other physiological measurements. The algorithm combined 27 of the most common measurements used to diagnose septic shock and generated a targeted real-time warning score, or TREWScore. If a patient’s TREWScore indicated that he or she was at risk for septic shock, an alert was sent to a doctor who could then take action when the sepsis was relatively easy to counteract.
The study found that TREWScore identified 61 percent of the septic shock patients before one of its competitors, Modified Early Warning Score did. Not only can TREWScore predict septic shock sooner, it is also more accurate. TREWScore correctly spotted 85 percent of the patients who were at risk for septic shock, while the Modified Early Warning Score correctly identified septic shock 73 percent of the time.
More than 750,000 people in the United States develop severe sepsis and septic shock each year; for 40 percent of them, the condition is ultimately fatal, says Hagar.
One hurdle that needs to be overcome is that electronic health data can be challenging to work with, explains the study’s lead computer engineer, Suchi Saria. Part of the difficulty is that there may be systematic bias in the medical information recorded. For instance, a patient who is successfully treated will appear as low-risk in the electronic health records because the septic shock was prevented. These cases hurt the algorithm’s performance. Current computerized clinical decision support (CDS) models that utilize electronic health records do not account for this kind of censored information.
Saria and her team address this problem by modifying pattern recognition algorithms so the computer can avoid mistaking high-risk patients for low-risk ones. This computer system can be tailored to many different medical conditions including acute lung injury, pneumonia, and post-rehabilitation illnesses like neuropathy.
“We are at a very exciting time,” says Saria. “More and more data is being collected on the electronic health records, and now our algorithms are reaching a point where they can be a real aid to clinicians.”