Microsoft Research Connections Blog

The Microsoft Research Connections blog shares stories of collaborations with computer scientists at academic and scientific institutions to advance technical innovations in computing, as well as related events, scholarships, and fellowships.

Software promises timely diagnosis of pneumonia in ICU patients

Software promises timely diagnosis of pneumonia in ICU patients

  • Comments 0

The advent of electronic medical records is transforming the care of patients in hospitals, clinics, and doctors’ offices around the world. Offering complete documentation of a patient’s medical history, digitized records are improving the accuracy of diagnoses and the continuity of patient care, which, in turn, means improved patient outcomes.

Electronic medical records can be especially useful in the diagnosis of pneumonia, which has a nasty habit of appearing after a patient has been hospitalized in the intensive care unit (ICU). Currently, such diagnoses are made by consensus, after a thorough chart review of the patient’s medical tests and clinical notes.

This means that a physician must comb through hand-written records and copious test results to reach the correct diagnosis—a time-consuming, resource-intensive process. Now, my colleagues and I at the University of Washington, in collaboration with Microsoft researcher Lucy Vanderwende and using Microsoft Research Splat (Statistical Parsing and Linguistic Analysis Toolkit), have created deCIPHER, a project that demonstrates the potential to use natural language processing (NLP) and machine learning to diagnose such critical illnesses automatically from the patient’s electronic medical records.

We studied the diagnosis of pneumonia in approximately 100 patients being treated in the ICU at Seattle’s Harborview Medical Center. By using the electronic medical records of these patients—whose pneumonia diagnoses had already been established by clinical consensus—we employed state-of-the-art NLP tools from Microsoft to identify the critical clinical information. We then ran that data through a machine-learning framework to see if the software could be trained to correctly diagnose the pneumonia cases based solely on an automated review of the digital medical records. The results were so promising—the software achieved a correct diagnosis with correct time-of-onset for positive cases in 84 percent of the patients—that our clinical collaborators are considering the addition of our pneumonia-detection models to the dashboard they use to monitor ICU patients.

Meliha Yetisgen, Assistant Professor of Biomedical Informatics, University of Washington

Related links

Leave a Comment
  • Please add 6 and 8 and type the answer here:
  • Post