ABSTRACT
Physicians collect data in patient encounters that they use to diagnose patients. This process can fail if the needed data is not collected or if physicians fail to interpret the data. Previous work in orofacial pain (OFP) has automated diagnosis from encounter notes and pre-encounter diagnoses questionnaires, however they do not address how variables are selected and how to scale the number of diagnoses. With a domain expert we extract a dataset of 451 cases from patient notes. We examine the performance of various machine learning (ML) approaches and compare with a simplified model that captures the diagnostic process followed by the expert. Our experiments show that the methods are adequate to making data-driven diagnoses predictions for 5 diagnoses and we discuss the lessons learned to scale the number of diagnoses and cases as to allow for an actual implementation in an OFP clinic.
