Paramedics respond to a 911 call to find an elderly patient with difficulty breathing. Is he suffering from acute emphysema or heart failure? The symptoms look the same, but initiating the wrong treatment could cause severe complications.
Researchers from MIT’s Research Laboratory of Electronics (RLE) and colleagues at Harvard Medical School and the Einstein Medical Center in Philadelphia believe that repurposing a piece of medical equipment that’s standard in all U.S. and European ambulances could help paramedics make this type of diagnosis.
They’ve developed an algorithm that can use readings from a capnograph—a machine that measures the concentration of carbon dioxide in a patient’s exhalations—to determine, with high accuracy, whether a patient is suffering from emphysema or heart failure.
“This machine is ubiquitous,” says professor of electrical and biomedical engineering George Verghese. “It’s actually in every emergency department and operating room. But the use that they’ve typically made of it is much more limited than what we were attempting here.”
Capnography is principally used to aid medical professionals as they insert breathing tubes into the tracheas of sedated patients. If the capnograph displays a regular wave pattern, with crests for exhalations and troughs for inhalations, the tube has been inserted properly. If the capnogram flatlines, the tube has mistakenly been inserted into the esophagus.
Over time, physicians observed that the capnograms of patients with congestive heart failure and emphysema were subtly but consistently different both from each other and from those of healthy subjects. So Verghese and Thomas Heldt, leaders of the Computational Physiology and Clinical Inference Group at RLE, recruited master’s student Rebecca Mieloszyk to investigate capnography’s possible diagnostic applications.
Mieloszyk’s approach was a somewhat unconventional twist on machine learning. First, she split capnogram data from 84 study subjects into 50 overlapping but nonidentical subsets. Then she turned an algorithm loose on the data to look for correlations between the patients’ capnograms and their ultimate diagnoses. The algorithm found slightly different patterns in each subset, generating 50 different rules. Mieloszyk then applied those rules to a separate set of 55 patient records, arriving at a diagnosis for each one by tallying the “votes” of the 50 rules.
Diagnostic techniques are assessed by graphing their true-positive rates against their false-positive rates and measuring the area under the resulting curve. A measurement of one is perfect. The researchers found that their algorithm for distinguishing healthy subjects from those with emphysema yielded an area under the curve of 0.98. The algorithm that distinguished emphysema patients from those with congestive heart failure checked in at 0.89.
Verghese and Heldt’s team is currently testing the algorithm with paramedics in the field. The researchers are also evaluating whether capnography can measure the severity of asthma attacks and the degree of sedation in patients undergoing medical procedures.