Analysis of Opioid Related Adverse Events and Signal Detection with Machine Learning
Open Access
- Author:
- Davaasuren, Dolzodmaa
- Graduate Program:
- Industrial Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- July 26, 2019
- Committee Members:
- Soundar Kumara, Thesis Advisor/Co-Advisor
Robert Voigt, Committee Member - Keywords:
- Opioid drugs
Adverse Drug Events
Adverse Drug Reactions
Drug-Drug Interaction - Abstract:
- In the past 30 years, the United States has experienced an unprecedented opioid epidemic. It’s been reported that physicians dispense more than 650,000 opioid prescriptions per average day, according to [1]. Opioids are the most popular choice of drugs of pain management due to its high efficacy. However, their usage is commonly associated with a variety of adverse drug events (ADEs), ranging from nausea and vomiting to urinary retention and respiratory depression. It’s been reported that nearly 10% of total Adverse Drug Events are associated with opioids [2]. Therefore, it is crucial to detect patterns in opioid-related ADEs based on patient information. In this thesis, we present a data-driven approach to detect any opioid-related serious outcomes given a limited amount of information about the patients’ medical history and demographic information. We are using the FDA’s Adverse Event Reporting database (FAERS). FAERS database contains the quarterly released reports made of individual components such as drug, patient demographics, drug indications, drug-related reactions, and resulting outcome data. First of all, we performed extensive data pre-processing on each individual dataset to make it a cleaner, normalized, and more informative. Secondly, we analyzed each dataset more in-depth with basic data mining methods to get preliminary results and insights about the opioid-related adverse events. Lastly, we used an unsupervised clustering method to detect most informative distinct clusters of reports in the pool and found out different types of most common observations in the database. Additionally, random tree classifiers and deep neural networks were used to predict a seriousness of outcome (DE–death) and serious drug reactions such as renal impairment, medication error, death, and congenital disorder and gastrointestinal disorder. Despite the small number of signals, these methods showed relatively high accuracy in detecting abnormal reactions. Also, we developed a way to measure the similarity between two reports, the one with known DDIs and the other without known DDIs. As a result, we found out potential undocumented DDIs using a drug to drug structure similarity and rate of common reactions among each report pair.