Date of Award

5-11-2025

Date Published

June 2025

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Social Science

Advisor(s)

Peng Gao

Keywords

artificial neural network;data imputation;emergency medical services;EMS;machine learning;population health

Subject Categories

Geography | Social and Behavioral Sciences

Abstract

Understanding health disparities is fundamental to population health and is essential for design and promotion of health policies that protect individuals from inequitable access, unfair treatment, and discrimination. However, despite broad reliance on technology and digital data, missingness is a common issue in patient health records. This research aspired to increase understanding of health outcome divergence through the lens of emergency medical events in the United States. Retrospective study focused on outcome missingness in the National Emergency Medical Services Information System (NEMSIS) years 2017 through 2023, a voluminous public research dataset consisting of roughly a quarter billion patient events. Only a small fraction (<1%) of NEMSIS records include a definitive end-of-event status i.e., the patient “lived” or “died.” As a result, while the U.S. 9-1-1 emergency medical response systems are comprehensive, gauging success of their principle function, safely delivering a live individual, who is experiencing a life-threatening medical event, to a hospital emergency department, is challenging. A consequence of outcome missingness is that when differences occur in health care for any reason – including social factors, local resource inequities, discrimination, etc. – impact on patient survival is untestable. U.S. history of EMS and its evolution to prehospital health care are important because more and more people use 9-1-1 services in lieu of access to primary care and for medical emergencies caused by untreated conditions. This equal access is protected by the Emergency Medical Treatment and Labor Act (“EMTALA”), which prevents hospitals and EMS agencies receiving Medicare funds from refusing to treat patients. For now, EMS is a social equalizer by its natural control of access. However, recent studies have shown localized EMS disparities by race/ethnic, sex, social, and geographic location. This dissertation research offers the following contributions. (i) Accurately imputed individual patient “dead” or “alive” outcomes are imputed for >150 million EMS patient events (2017 to 2023). Machine learning methodology trained, tested, and cross-validated a multi-layer neural network model that improved “dead” category accuracy (sensitivity; recall) by 35 percentage points over a previously published method. The model achieved almost 90% balanced accuracy for both “dead” and “alive” predictions despite severe category imbalance, and without evidence of over-fitting. Stratified K-fold cross validation was performed to minimize general bias and to provide sensitivity analysis best-practice to address missingness. (ii) First-of-a-kind national EMS mortality benchmarks are estimated for out-of-hospital cardiac arrest (OHCA), opioid overdose, respiratory arrest, sepsis, ST-segment elevation myocardial infarction (STEMI), stroke, trauma, and unresponsive patient. The benchmarks provide reference points for future research to evaluate interventions and sub-population outcome disparities. (iii) Several disparate mortality patterns are revealed by medical emergency category, sex, race/ethnicity, age group, and geographic location. (iv) Models and systems for responsible AI-based population health are advanced by rigorously demonstrating that machine learning accurately predicts missing EMS patient outcomes. (v) New methodology is demonstrated for analyzing and visualizing threshold selection and specificity / sensitivity trade-off in binary classification, and for estimating population-level mortality rates by leveraging fractional response regression. (vi) Results from a previously published study are reproduced and replicated to anchor the work by comparison to the previous authors’ hand-crafted imputed outcomes. (vii) Finally, this work contributes to the understanding of population health and the notion of missing data as a socio-technical determinant of health.

Access

Open Access

Download

Included in

Geography Commons

Share

COinS