Contextualizing Medical Image Analyses with Electronic Health Histories

Abstract

Diseases of the optic nerve affect millions of Americans each year. These diseases include a wide range of conditions such as glaucoma, thyroid eye disease, optic neuritis, papilledema, idiopathic intracranial hypertension, and orbital infections. The optic nerve is a very delicate structure and often, the window for intervention is very small. Diagnosis of orbital conditions is done based on visual testing, careful review of the patient’s history, and orbital imaging. In this body of work, we present data-driven methods to develop a system that performs automated analysis of a patient’s electronic medical history including past clinical visits and imaging, to identify salient features and develop models for predicting clinical outcomes in optic nerve disease. As a first step in this process, image processing methods were developed to segment the relevant structures in the eye orbit, including the optic nerve, the globe, the extraocular muscles and orbital fat, and extract structural measurements. Next, it was shown that the measurements extracted from orbital imaging are correlated with visual function, and the phenotypes extracted from our statistical models match with clinical findings of disease subtypes. As a part of this research, an open-source tool was developed to extract clinical phenotypes from other electronic medical record data such as past diagnoses and procedures, which were used to develop predictive models for disease diagnosis as well. The main contribution of this dissertation was to show that context-aware models that integrate imaging biomarkers along with other EMR features extracted from a patient’s electronic health history significantly improved outcome modelling when compared to traditional image analysis studies.