Cardiovascular disease (CVD) affects millions of people and is a leading cause of death worldwide. CVD consists of a broad set of conditions including structural heart disease, coronary artery disease and stroke. Risk for each of these conditions accumulates over long periods of time depending on several risk factors. In order to reduce morbidity and mortality due to CVD, preventative treatments administered prior to first CVD event are critical. According to clinical guidelines, such treatments should be guided by an individual’s total risk within a window of time. A related objective is secondary prevention, or early detection, wherein the aim is to identify and mitigate the impact of a disease that has already taken effect. With the widespread adoption of electronic health records (EHRs), there is tremendous opportunity to build better methods for risk assessment and early detection. However, existing methods which use EHRs are limited in several ways: (1) they do not leverage the full longitudinal history of patients, (2) they use a limited feature set or specific data modalities, and (3) they are rarely validated in broader populations and across different institutions. In this dissertation, I address each of these limitations. In Aim 1, I explore the challenge of handling longitudinal, irregularly sampled clinical data, proposing discriminative and generative approaches to model this data. In Aim 2, I develop a multimodal approach for the early detection of structural heart disease. Finally, in Aim 3, I study how different feature inclusion choices affect the transportability of deep risk assessment models of coronary artery disease across institutions. Collectively, this dissertation contributes important insights towards building better approaches for risk assessment and early detection of CVD using EHR data and systematically assessing their transportability across institutions and populations.

No description available.

<p>We evaluated the comparability of influenza-related events self-reported by research participants and their outpatient medical records using data collected from the Hutterite Influenza Prevention Trial. We also explored the implications of using data on influenza symptoms from both data sources, independently and in combination, as predictors of laboratory-confirmed influenza. Self-report of influenza symptoms, physician-diagnosed otitis media and antibiotics prescribed at outpatient consultations was collected from trial participants. Similar data were also collected by fax requests for medical record information to the medical facilities. We found lower rates of self-reported prevalence for fever, sore throat, earache and otitis media and higher rates of antibiotic prescriptions compared to the medical records. Total agreements between self-report and medical report of symptoms varied between 61% and 88%. Negative agreement was considerably higher than positive agreement for each symptom, except cough. Self report of otitis media was a very specific measure (93%), but had lower sensitivity (47%). Positive predictive value was moderate at 64% but negative predictive value was good at 86%. Self-reported antibiotic prescription was a highly sensitive measure (98%), but had low specificity (50%). Positive predictive value was high at 91% but negative predictive value was modest at 65%. Fever (on its own) and combined with cough and/or sore throat were highly correlated with laboratory-confirmed influenza for all data sources. The ILI surveillance definition of fever and sore throat, based on combined symptoms by both medical records and self report, was the best predictor laboratory confirmed influenza.</p> / Doctor of Philosophy (PhD)

Medical records are rapidly being digitized to electronic medical records. Although Electronic Medical Records (EMRs) improve administration, billing, and logistics, an open research problem remains as to how doctors can leverage EMRs to enhance patient care. This thesis describes a system that analyzes a patient's evolving EMR in context with available biomedical knowledge and the accumulated experience recorded in various text sources including the EMRs of other patients. The aim of the Clinician Decision Support (CDS) Dashboard is to provide interactive, automated, actionable EMR text-mining tools that help improve both the patient and clinical care staff experience. The CDS Dashboard, in a secure network, helps physicians find de-identified electronic medical records similar to their patient's medical record thereby aiding them in diagnosis, treatment, prognosis and outcomes. It is of particular value in cases involving complex disorders, and also allows physicians to explore relevant medical literature, recent research findings, clinical trials and medical cases. A pilot study done with medical students at the Virginia Tech Carilion School of Medicine and Research Institute (VTC) showed that 89% of them found the CDS Dashboard to be useful in aiding patient care for doctors and 81% of them found it useful for aiding medical students pedagogically. Additionally, over 81% of the medical students found the tool user friendly. The CDS Dashboard is constructed using a multidisciplinary approach including: computer science, medicine, biomedical research, and human-machine interfacing. Our multidisciplinary approach combined with the high usability scores obtained from VTC indicated the CDS Dashboard has a high potential value to clinicians and medical students. / Master of Science

The rapid adoption of Electronic Health Records (EHRs)--electronic versions of a patient's medical history--has been instrumental in streamlining administrative tasks, increasing transparency, and enabling continuity of care across providers. An unintended consequence of the increased documentation burden, however, has been reduced face-time with patients and, concomitantly, a dramatic rise in clinician burnout. Time spent maintaining and making sense of a patient's electronic record is a leading cause of burnout. In this thesis, we pinpoint a particularly time-intensive, yet critical, documentation task: generating a summary of a patient's hospital admissions, and propose and evaluate automated solutions. In particular, we focus on faithfulness, i.e., accurately representing the patient record, and completeness, i.e., representing the full context, as the sine qua non for safe deployment of a hospital-course summarization tool in a clinical setting. The bulk of this thesis is broken up into four chapters: §2 Creating and Analyzing the Data, §3 Improving the Faithfulness of Summaries, §4 Measuring the Faithfulness of Summaries, and, finally, §5 Generating Grounded, Complete Summaries with LLMs. Each chapter links back to the core themes of faithfulness and completeness, while the chapters are linked to each other in that the findings from each chapter shape the direction of subsequent chapters. Given the documentation authored throughout a patient's hospitalization, hospital-course summarization requires generating a lengthy paragraph that tells the story of the patient admission. In § 2, we construct a dataset based on 109,000 hospitalizations (2M source notes) and perform exploratory analyses to motivate future work on modeling and evaluation [NAACL 2021]. The presence of highly abstractive, entity dense references, coupled with the high stakes nature of text generation in a clinical setting, motivates us to focus on faithfulness and adequate coverage of salient medical entities. In § 3, we address faithfulness from a modeling perspective by revising noisy references [EMNLP 2022] and, to reduce the reliance on references, directly calibrating model outputs to metrics [ACL 2023]. These works relied heavily on automatic metrics as human annotations were limited. To fill this gap, in §4, we conduct a fine-grained expert annotation of system errors in order to meta-evaluate existing metrics and better understand task-specific issues of domain adaptation and source-summary alignments. We find that automatically generated summaries can exhibit many errors, including incorrect claims and critical omissions, despite being highly extractive. These errors are missed by existing metrics. To learn a metric which is less correlated to extractiveness (copy-and-paste), we derive noisy faithfulness labels from an ensemble of existing metrics and train a faithfulness classifier on these pseudo labels [MLHC 2023]. Finally, in § 5, we demonstrate that fine-tuned LLMs (Mistral and Zephyr) are highly prone to entity hallucinations and cover fewer salient entities. We improve both coverage and faithfulness by performing sentence-level entity planning based on a set of pre-computed salient entities from the source text, which extends our work on entity-guided news summarization ([ACL, 2023] and [EMNLP, 2023]).

Understanding the genetic basis of complex disease is a critical research goal due to the immense, worldwide burden of these diseases. Observational data, such as electronic health records (EHR), offer numerous advantages in the study of complex disease genetics. These include their large scale, cost-effectiveness, information on many different conditions, and future scalability with the widespread adoption of EHRs. Observational data, however, are challenging for research due to noise and confounding. EHR data reflect factors including the healthcare process and access to care, as well as broader societal effects like systemic biases. Billing codes for complex diseases may be recorded when no diagnosis is intended, and they may be missing when a diagnosis would be correct. Overall, systematic errors distort the genetic signal available for study and motivate taking a closer look at the ways that phenotypes can be defined using observational data. In Chapter 3, we introduce MaxGCP, a novel phenotyping method designed to purify the genetic signal in observational data. Our approach optimizes a phenotype definition to maximize its coheritability with the complex trait of interest. We first validated this method in simulations of 5000 different phenotypes across a wide range of simulation parameters, demonstrating that the method improves genome-wide association study (GWAS) power compared to conventional methods. Having evaluated it in simulation, we next applied the method in real data analyses of stroke and Alzheimer’s disease. By comparing GWAS associations to high-quality, independent test data, we were able to compare both the sensitivity and specificity of our method. This analysis similarly found that MaxGCP boosts GWAS power compared to previous methods. In Chapter 4, we extend this work to increase the speed and re-usability of pan-biobank GWAS with another new method, Indirect GWAS. Large scale, pan-biobank studies provide a powerful resource in complex disease genetics, generating shareable summary statistics on thousands of phenotypes. Biobank-scale GWAS have two notable limitations: they are resource-intensive to compute and do not inform about hand-crafted phenotype definitions, which are often more relevant to study. Our method uses summary statistics to addresses these limitations. It computes GWAS summary statistics for any phenotype defined as a linear combination of other phenotypes. We demonstrate a number of useful applications, including an order of magnitude improvement in runtime for large-pan-biobank GWAS and ultra-rapid (less than one minute) GWAS on hand-crafted phenotype definitions using only summary statistics. Through the development of new computational and statistical methods, this work demonstrates the importance and power of the phenotype side of genetic association studies, and it provides two new approaches that can improve future genetic studies of complex disease.

No description available.

The purpose of this project was to develop a document imaging application. By scanning the documents into an electronic repository, medical staff will be able to more easily store and locate these records. To make the application user friendly and facilitate staff access to patient medical records, the application is wed-based and uses the Oracle Application Server to implement a multitiered model.

published_or_final_version / Community Medicine / Doctoral / Doctor of Philosophy

Information Systems (IS) innovations in healthcare sector are seen as panacea to control burgeoning demand on healthcare resources and lack of streamlining in care delivery. Two particular manifestations of such innovations are telehealth and electronic records in its two forms: the electronic medical records and the electronic health records. Deployment efforts concerning both of these IS-innovations have encountered a rough terrain and have been slow. Problems are also faced while evaluating the effectiveness of innovations on health and care delivery outcomes through strategies such as randomised controlled trials- particularly in case of telehealth. By taking these issues into account, this research investigates the issues that affect IS innovation deployment and its evaluation. The strategy adopted in this research was informed by recursive philosophy and theoretical perspectives within IS that strived to expound this recursive relationship. It involved conducting two longitudinal case studies that are qualitative in nature. The first study involved telehealth deployment and its evaluation in the UK, while the second case study involved the deployment of electronic medical/health records in the US. Data was collected through focus group discussions, interviews and online discussion threads; and was analysed thematically. The results of this research indicate that there are nine issues that arise and affect the deployment and evaluation of IS innovation in healthcare; and these are design, efficiency and effectiveness, optimality and equity, legitimacy, acceptance, demand and efficacy, expertise, new interaction patterns, and trust. These issues are attributes of relationships between the IS innovation, context of healthcare and the user. The significance of these attributes varies during the deployment and evaluation process, and due to iterative nature of IS innovation. This research further indicates that all the attributes have either direct or indirect impact on work practices of the user.