Computerized automated pulmonary disease diagnosis

This thesis details the design and implementation of a computerized system which generates a list of 5 differential diagnoses on patients with suspected pulmonary disease. A probabilistic diagnostic model, sequential applications of Bayes' Theorem, was used to calculate a series of likelihoods that a patient would have each of 29 different pulmonary diseases. This approach to automated medical diagnosis was dependent on the facilities provided by the HELP system at LDS Hospital, a 550 bed tertiary care hospital located in Salt Lake City, Utah. After development and debugging of the medical logic, a retrospective study on 69 hospitalized patients with previously diagnosed pulmonary disease was conducted to measure the success of this system. The predictive logic was used to calculate the 5 most likely diseases for each patient. This differential diagnosis list contained the appropriate discharge diagnosis in 55 of 69 patients (80%). The medical logic (total 29 disease modules) yielded a sensitivity of 80% and specificity of 97%. In a study of a subgroup of 18 patients having multiple pulmonary diseases the system correctly classified all pulmonary diseases present in 33% (6/18) of the cases. The overall sensitivity of this subgroup of patients with multiple pulmonary diseases was 64% and the specificity was 91%. As judged from this preliminary study of the diagnostic logic, the method proved remarkably robust. These promising results have led to the installation of this system in a clinical setting. It will be used in a prospective manner to study patients who have chest radiographs performed in the Radiology Department at LDS Hospital.