The increasing adoption of electronic health records (EHRs) has led to an unprecedented amount of patient health information stored in an electronic format. The ability to comb through this information is imperative, both for patient care and computational modeling. Creating a system to minimize unnecessary EHR data, automatically distill longitudinal patient information, and highlight salient parts of a patient’s record is currently an unmet need. However, summarization of EHR data is not a trivial task, as there exist many challenges with reasoning over this data. EHR data elements are most often obtained at irregular intervals as patients are more likely to receive medical care when they are ill, than when they are healthy. The presence of narrative documentation adds another layer of complexity as the notes are riddled with over-sampled text, often caused by the frequent copy-and-pasting during the documentation process. This dissertation synthesizes a set of challenges for automated EHR summarization identified in the literature and presents an array of methods for dealing with some of these challenges. We used hybrid data-driven and knowledge-based approaches to examine abundant redundancy in clinical narrative text, a data-driven approach to identify and mitigate biases in laboratory testing patterns with implications for using clinical data for research, and a probabilistic modeling approach to automatically summarize patient records and learn computational models of disease with heterogeneous data types. The dissertation also demonstrates two applications of the developed methods to important clinical questions: the questions of laboratory test overutilization and cohort selection from EHR data.

Complex patient are becoming more and more of a challenge to the health care system given the amount of care they require and the amount of documentation needed to keep track of their state of health and treatment. Record keeping using the EHR makes this easier but mounting amounts of patient data also means that clinicians are faced with information overload. Information overload has been shown to have deleterious effects on care, with increased safety concerns due to missed information. Patient record summarization has been a promising mitigator for information overload. Subsequently, a lot of research has been dedicated to record summarization since the introduction of EHRs. In this dissertation we examine whether unsupervised inference methods can derive patient problem-oriented summaries, that are robust to different patients. By grounding our experiments with HIV patients we leverage the data of a group of patients that are similar in that they share one common disease (HIV) but also exhibit complex histories of diverse comorbidities. Using a user-centered, iterative design process, we design an interactive, longitudinal patient record summarization tool, that leverages automated inferences about the patient's problems. We find that unsupervised, joint learning of problems using correlated topic models, adapted to handle the multiple data types (structured and unstructured) of the EHR, is successful in identifying the salient problems of complex patients. Utilizing interactive visualization that exposes inference results to users enables them to make sense of a patient's problems over time and to answer questions about a patient more accurately and faster than using the EHR alone.

A patient monitor device for use during open heart surgery has been designed and constructed. The device uses a VIC 20 microcomputer along with some additional circuitry to monitor 3 separate functions. The first patient variable monitored is the blood flow rate through the extracorporeal blood circuit during surgery. The device also continuously monitors and displays 6 separate temperatures. Finally, 3 individual timers are monitored and displayed with the device. Both the hardware and the software used in the design are fully described. / Master of Science