Neural network analysis of the volumetric capnogram to detect pulmonary embolism

Background: Pulmonary embolism (PE) produces ventilation/perfusion mismatch that may be manifested in various variables of the volume-based capnogram (VBC). We hypothesized that a neural network (NN) system could detect chang es in VBC variables that reflect the presence of a PE. Methods: A commercial VBC system was used to record multiple respiratory va riables from consecutive expiratory breaths. Data from 12 subjects (n = 6 P E+ and n = 6 PE-) were used as input to a fully connected back-propagating NN for model development. The derived model was tested in a prospective, ob servational study at an urban teaching hospital. Volumetric capnograms were then collected on 53 test subjects: 30 subjects with PE confirmed by pulmo nary angiography or diagnostic scintillation lung scan, and 23 subjects wit hout PE based on pulmonary angiography. The derived NN model was applied to VBC data from the test population. Results: Seventeen VBC variables were used by the derived NN model to gener ate a numeric probability of PE. When the derived NN model was applied to V BC data from the 53 test subjects, PE was detected with a sensitivity of 10 0% (95% CI = 89% to 100%) and a specificity of 48% (95% CI = 21% to 69%). T he likelihood ratio positive [LR(+)] for the VBC-NN test was 1.82 and the L R (-) was 0.1. Conclusion: This study demonstrates the feasibility of developing a rapid, noninvasive breath test for diagnosing PE using volumetric capnography and NN analysis.