Acute respiratory distress syndrome (ARDS) is a severe inflammatory lung disease with high mortality risk. Development of new and effective therapies for ARDS has been slow due to a lack of precision in its diagnostic criteria. We report preliminary research to extract computational and semantic features directly from chest X-ray images that are used to train machine learning classifiers. Our approach demonstrates the feasibility of using machine learning to identify radiographic criteria that are more consistent and accurate for the diagnosis of ARDS.