Temporal analysis of emergency department frequent users using machine learning approaches

With a wealth of electronic health records (EHRs) available, there is an opportunity to better utilize the existing healthcare infrastructure by anticipating patient utilization to reduce healthcare costs. Emergency Department (ED) is one such expensive healthcare facility. However, emergency services are disproportionately used by a small set of users known as frequent ED users (EDFU) also called frequent flyers. They tend to be heavy users of other parts of the health care system and unable to get the appropriate care they need. By identifying these patients, programs can be developed to improve their health outcomes and reduce overall costs. In our study, we develop a prediction model which identifies frequent ED users based on their past electronic health records (EHR). We make use of regional trauma level 1 hospital data from 2011-2015, to predict whether the patient is an EDFU. We try to do that by leveraging varied pieces of information from the patient electronic health record for a year and predicting their behavior for the next year. We propose an end-to-end deep learning framework that uses past EHR to classify patients into frequent and non-frequent/occasional users. A diagnosis embedding inspired by the Skip-gram-based word2vec embedding framework was developed as input to the EDFU prediction model. We find that logistic regression models perform comparably well than recurrent neural networks and gradient-boosting machines. Embeddings perform slightly better than one-hot encoded diagnosis codes for both the logistic regression model and gradient boosting machine. However, presence of embeddings do not have any effect on recurrent neural networks. The second goal of this study is to investigate frequent users by analyzing their primary diagnosis codes for the first four visits. We define distance between any two visits based on their primary diagnosis. We do that by making use of similarity measures developed for semantic trees in natural language processing (NLP). We make use of the same dataset as before and isolate visits when the patient first became a frequent user. A proximity score is then defined which is the average of distances between all possible combinations of the first four visits when the patient first became a frequent user. The patients are then clustered into three distinct classes based on their proximity score. On analyzing these clusters for various visit and patient parameters, we find that patients with similar visits tend to be a higher frequency of visits than patients with dissimilar visits. Patients in the first cluster are discharged more often and arrive by Walk-In than other modes. We further elaborate on the diagnosis codes that characterize the patients in each of these three clusters. Such an analysis of patients can help EDs find patterns of diagnosis codes and help them develop strategies to address patients who come to ED with similar diagnoses.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Industrial Engineering

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2022

Author dc:creator

• Fernandes, Lloyd

Contributors dc:contributor

• Kang, Hyojung
• Sreenivas, Ramavarapu S

• Emergency Department
• Healthcare Machine Learning
• Machine Learning
• Ehr

Statement dc:rights

• Copyright 2022 Lloyd Fernandes

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/117829

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F117829