Date of Award

January 2020

Document Type

Thesis

Degree Name

Master of Public Health (MPH)

Department

School of Public Health

First Advisor

Daniel Weinberger

Second Advisor

David Paltiel

Abstract

Introduction: Legionnaires’ disease (LD), caused by the Legionella bacterium, is a rare but serious atypical pneumonia that is often underdiagnosed in the clinical setting. Recent guideline changes by the American Thoracic Society limited use of the urinary Legionella tests, the primary form of diagnosis, due to the excessive costs of regular testing. Delays in testing and diagnosis of LD result in increased cost of treatment and risk of death.

Objective: To develop a predictive model to estimate a patient’s probability of having LD. This model would serve as a screen to better target diagnostic efforts, prevent misdiagnosis, and reduce the overall cost of treatment for patients with LD.

Methods: Commonly collected hospital admission data including age, sex, admission month, and clinical diagnosis were used to create backward elimination logistic regression models. Additional non-hospital variables included smoothed incidence in home zip code and weather. Models were trained on four data sets of community acquired pneumonia cases that differed by location: New Jersey, New York (excluding NYC), New York City, and a combined data set of all locations. A decision analysis was used to determine a cost-minimizing threshold.

Results: Using the models as a screen to guide diagnostic testing produced a wide range of sensitivities despite consistently high specificities. The model trained on New York data consistently performed the best on out of sample validation with the sensitivity of the screen ranging between 0.96 and 0.99. Additionally, sensitivity analyses demonstrated that across a range of prevalences, the models resulted in a lower cost per patient compared to testing all patients. As the proportion of Legionella-attributable hospitalized pneumonia cases increased to the expected proportion, the average savings per patient increased as well.

Conclusion: The models presented here can help guide clinicians in determining who should or should not be tested for LD with the Legionella urine antigen test to reduce underdiagnosis, decrease time to diagnosis, and improve health outcomes.

Comments

This thesis is restricted to Yale network users only. It will be made publicly available on 05/27/2021

Share

COinS