"I tried to organize the protocol into a flow chart, but it was a mess," said Martin. "It's not easy to convert medical protocols into ones and zeros because there is a lot of nuance and judgment involved."
After a period of "beating our heads against the wall trying to make this work," the ISIS team suggested a new approach, something they called a "state-based decision engine." Essentially, this involved breaking down the guidelines into a series of independent processes that can take place sequentially or simultaneously. "This really captures the way doctors work. If we see low blood pressure, then we think of one set of treatments. If we see low blood sugar, then we think of another set. If we see the two together, then we consider a third set of possible measures we can take," Martin said. "The more we kicked it around, the more we liked it."
Using this approach, Mathe and his colleagues developed a special modeling language specifically for clinical decision-making. "Although the language is specific to sepsis management, we made the underlying technical infrastructure so general that it can model virtually any medical protocol," Mathe said. The team has already begun applying it to a second problem, treatment of chronic heart failure.
According to the researchers, it will take six to 12 months of operation before they can begin to judge the system's effectiveness. They began running the detection system in the background in March. During this time they averaged three to four alerts per day in the 25-bed ICU. When the system was deployed, the very first alert led an attending physician to begin antibiotic therapy. They will run the detection system for several months before imple
|Contact: David F. Salisbury|