Date of Award

Spring 2022

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Investigative Medicine

First Advisor

Krishnaswamy, Smita

Abstract

Acute gastrointestinal bleeding is a common and costly condition, accounting for over 2.2 million hospital days and 19.2 billion dollars of medical charges annually. Risk stratification is a critical part of initial assessment of patients with acute gastrointestinal bleeding. Although all national and international guidelines recommend the use of risk-assessment scoring systems, they are not commonly used in practice, have sub-optimal performance, may be applied incorrectly, and are not easily updated. With the advent of widespread electronic health record adoption, longitudinal clinical data captured during the clinical encounter is now available. However, this data is often noisy, sparse, and heterogeneous. Unsupervised machine learning algorithms may be able to identify structure within electronic health record data while accounting for key issues with the data generation process: measurements missing-not-at-random and information captured in unstructured clinical note text. Deep learning tools can create electronic health record-based models that perform better than clinical risk scores for gastrointestinal bleeding and are well-suited for learning from new data. Furthermore, these models can be used to predict risk trajectories over time, leveraging the longitudinal nature of the electronic health record. The foundation of creating relevant tools is the definition of a relevant outcome measure; in acute gastrointestinal bleeding, a composite outcome of red blood cell transfusion, hemostatic intervention, and all-cause 30-day mortality is a relevant, actionable outcome that reflects the need for hospital-based intervention. However, epidemiological trends may affect the relevance and effectiveness of the outcome measure when applied across multiple settings and patient populations. Understanding the trends in practice, potential areas of disparities, and value proposition for using risk stratification in patients presenting to the Emergency Department with acute gastrointestinal bleeding is important in understanding how to best implement a robust, generalizable risk stratification tool. Key findings include a decrease in the rate of red blood cell transfusion since 2014 and disparities in access to upper endoscopy for patients with upper gastrointestinal bleeding by race/ethnicity across urban and rural hospitals. Projected accumulated savings of consistent implementation of risk stratification tools for upper gastrointestinal bleeding total approximately $1 billion 5 years after implementation. Most current risk scores were designed for use based on the location of the bleeding source: upper or lower gastrointestinal tract. However, the location of the bleeding source is not always clear at presentation. I develop and validate electronic health record based deep learning and machine learning tools for patients presenting with symptoms of acute gastrointestinal bleeding (e.g., hematemesis, melena, hematochezia), which is more relevant and useful in clinical practice. I show that they outperform leading clinical risk scores for upper and lower gastrointestinal bleeding, the Glasgow Blatchford Score and the Oakland score. While the best performing gradient boosted decision tree model has equivalent overall performance to the fully connected feedforward neural network model, at the very low risk threshold of 99% sensitivity the deep learning model identifies more very low risk patients. Using another deep learning model that can model longitudinal risk, the long-short-term memory recurrent neural network, need for transfusion of red blood cells can be predicted at every 4-hour interval in the first 24 hours of intensive care unit stay for high risk patients with acute gastrointestinal bleeding. Finally, for implementation it is important to find patients with symptoms of acute gastrointestinal bleeding in real time and characterize patients by risk using available data in the electronic health record. A decision rule-based electronic health record phenotype has equivalent performance as measured by positive predictive value compared to deep learning and natural language processing-based models, and after live implementation appears to have increased the use of the Acute Gastrointestinal Bleeding Clinical Care pathway. Patients with acute gastrointestinal bleeding but with other groups of disease concepts can be differentiated by directly mapping unstructured clinical text to a common ontology and treating the vector of concepts as signals on a knowledge graph; these patients can be differentiated using unbalanced diffusion earth mover’s distances on the graph. For electronic health record data with data missing not at random, MURAL, an unsupervised random forest-based method, handles data with missing values and generates visualizations that characterize patients with gastrointestinal bleeding. This thesis forms a basis for understanding the potential for machine learning and deep learning tools to characterize risk for patients with acute gastrointestinal bleeding. In the future, these tools may be critical in implementing integrated risk assessment to keep low risk patients out of the hospital and guide resuscitation and timely endoscopic procedures for patients at higher risk for clinical decompensation.

Share

COinS