Automated Knowledge Engineering Methods to Improve Consumers' Comprehension of their Health Records

PROJECT SUMMARY Today, more patients can access their health records online than ever before. However, clinical acronyms hinder patients' comprehension of their records and decrease the benefits of transparency. An automated system for expanding clinical acronyms should have major clinical significance and far-reaching consequences for improving patient-provider communication, shared decision-making, and health outcomes. Existing systems have limited power to expand clinical acronyms, primarily due to the lack of comprehensiveness (or generali- zability) of existing acronym sense inventories. Because developing comprehensive sense inventories is difficult, existing knowledge engineering methods have primarily focused on developing institution-specific sense inventories. Institution-specific sense inventories may not be generalizable to other geographical regions and medical specialties. Furthermore, developing an institution-specific sense inventory at every US healthcare organization is not feasible, especially without automated methods which currently do not exist. I developed advanced knowledge engineering methods to overcome these limitations through the use of fully automated techniques to generalize existing sense inventories from different geographical regions and medical specialties. My methods leverage the extensive resources already devoted to developing institution- specific sense inventories in the U.S., and may help generalize existing sense inventories to institutions without the resources to develop them. Although promising, challenges remain with the optimization and evaluation of these methods. The objective of the proposed project is to use knowledge engineering to improve patients' comprehension of their health records, focusing specifically on clinical acronyms. In Aim 1, I will develop new knowledge engineering methods to facilitate the automated integration of sense inventories, using literature- based quality heuristics and a Siamese neural network to establish synonymy. I will evaluate these methods using multiple metrics to assess redundancy, quality, and coverage in two test corpora with over 17 million clinical notes. In Aim 2, I will evaluate whether the knowledge engineering methods improve comprehension of doctors' notes in 60 hospitalized patients with advanced heart failure. With success, I will create novel, automated knowledge engineering methods that can be directly applied to improve patient care. This research is in support of my mentored doctoral training at Columbia University Department of Biomedical Informatics (DBMI) under Drs. David Vawdrey, George Hripcsak, Carol Friedman, Suzanne Bakken, and Chunhua Weng, and will include coursework on deep learning, oral presentations at major annual conferences, and career development planning, among other activities. DBMI is frequently recognized as one of the oldest and best programs of its kind in the world, and provides an exception training environment for my development into an independent and productive academic investigator.