Abstracts Category : Other

Design and validation of mobile kit and machinelearning algorithms for pulmonary disease screening anddiagnosis

by Daniel Chamberlain

Pulmonary diseases are responsible for more than 15%of deaths worldwide. Much of this burden is concentrated in thedeveloping world, where these diseases cause 19% of deaths. In muchof the developing world, pulmonary diseases are under-diagnosed andmisdiagnosed because the correct equipment is not available orhealth care is provided by workers with insufficient training. Tohelp improve the diagnosis of pulmonary disease, we built apulmonary diagnostic kit that consists of an electronicstethoscope, an augmented reality peak flow meter, and anelectronic questionnaire. Using this kit, we collected data frompatients who visited the Chest Research Foundation, a pulmonaryclinic in Pune, India. Using the data collected from thesepatients, we pursued several of avenues of research. First, wetrained algorithms to automatically detect two adventitious breathsounds: wheezes and crackles. We used two approaches to detectthese sounds: traditional signal processing methods and newtechniques from deep semi-supervised learning. Both techniquesshowed moderate success at identifying wheezes and crackles.Second, we evaluated the diagnostic potential of detecting wheezesand crackles and compared it to using signal processing analysis oflung sounds to directly detect pulmonary disease. We showed thatthis new technique leads to improved diagnostic accuracy. Thisfinding indicates that future research should focus less on lungsound identification. Third, we combined measurements from allthree components of our kit to predict the diagnosis of patientswith pulmonary disease. We showed that most of the diagnosticaccuracy of the kit was provided by the peak flow meter andquestionnaire combination. Together, these two devices were able toaccurately detect patients with asthma and COPD. After developingthe diagnostic algorithms, we built an Android application to guidea user through the necessary data collection to arrive at adiagnosis. The application was designed to create questionnairesand data queries from an externally defined model definition file,allowing the application to be easily repurposed for differentclassification tasks in medicine and other fields. Future researchwill expand the use of the pulmonary diagnostic kit to includeadditional pulmonary diseases and will test its use in alarge-scale field study to determine its accuracy as a screeningtool for asthma and COPD. If the results of future trials areconsistent with the findings in this thesis, the kit and algorithmcombination may provide useful information for improving diagnosisof pulmonary disease.Advisors/Committee Members: Richard R. Fletcher, Szolovits (advisor).