Suchi Saria, right, with first author Katherine Henry (Will Kirk, Homewood Photo, Johns Hopkins University)
6 August 2015. Researchers at Johns Hopkins University wrote a computer model that gives clinicians an early and accurate warning that a patient is developing sepsis, a life-threatening complication of infections. The team of medical researchers, computer scientists, and mathematicians published its findings yesterday in the journal Science Translational Medicine (paid subscription required).
Sepsis results from an immune-system reaction to chemicals released by the body to fight infection, including infections from medical equipment such as catheters. The inflammatory responses can occur anywhere in the body and generate a series of further reactions, including blood clots and leaking blood vessels, causing organ damage and failure. If sepsis develops into septic shock, blood pressure drops sharply, often causing death. Centers for Disease Control and Prevention says the number of people in U.S. hospitals with sepsis rose from 621,000 in 2000 to more than 1.1 million in 2008, with death resulting in 28 to 50 percent of cases.
The Johns Hopkins team, led by computer scientist and health policy professor Suchi Saria, is seeking tools for identifying potential sepsis cases earlier in their development, before they progress to septic shock. Current diagnostic methods, say the authors, can spot sepsis only in advanced stages, when there is little time for clinicians to respond. Current tools also cannot predict if sepsis cases will progress to septic shock.
Saria and colleagues identified 27 factors from a pool of 54 candidates that reliably predict sepsis development, and combined those factors into a machine-learning algorithm. The team then ran the algorithm with a database of more than 13,000 electronic medical records (without identifying information) of intensive-care patients at Beth Israel Deaconess Medical Center in Boston from 2001 to 2007 to refine the model, which they call Targeted Real-time Early Warning Score, or TREWScore. The electronic records include data captured from real-time vital sign monitors and lab reports.
The researchers tested the model with another 3,053 patient records from the same database, and compared the results with current tools for identifying risk of sepsis. The results show the TREWScore model performed with greater statistical certainty — identifying cases that develop into septic shock versus those that do not — than the current method, Modified Early Warning Score. TREWScore identified septic shock cases in advance by a median of 28 hours, with a sensitivity, or true positive rate, of 85 percent. In addition, TREWScore identified 59 percent more patients before signs of organ failure were noticed.
The researchers still need to devise workable processes for building the model into the systems and procedures of health care providers. “Our methods are reaching a point where they can be a real aid to clinicians,” says Saria in a university statement, “especially in noticing subtle hints, buried deep in a chart, that a problem is developing.” Saria notes, however, “we have to do this in a way that it is well-integrated into the existing clinical workflow and does not cause alarm fatigue.”
The authors note that with the Affordable Care Act, more opportunities will arise for constructing real-time predictive models like TREWScore. The Affordable Care Act requires health plans and providers to standardize billing and adopt electronic health records for cutting costs and reducing medical errors.
Read more:
- Computer Model Predicts Protein Binding to DNA, RNA
- Cough Diagnostics Mobile App in Clinical Trial
- Collaboration Collecting Data to ID Precise Cancer Meds
- Algorithms Compute Blood Volume from Video Images
- Pittsburgh Alliance Applies Big Data to Health Innovations
* * *
RSS - Posts
You must be logged in to post a comment.