Detecting adverse drug reactions in the general practice healthcare database
Publication Date
July 15, 2014
Creator
Abstract
The novel contribution of this research is the development of a supervised algorithm that extracts relevant attributes from The Health Improvement Network database to detect prescription side effects. Prescription drug side effects are a common cause of morbidity throughout the world. Methods that aim to detect side effects have historically been limited due to the data available, but some of these limitations may be overcome by incorporating longitudinal observational databases into pharmacovigilance. Existing side effect detecting methods using longitudinal observational databases have shown promise at becoming a fundamental component of post marketing surveillance but unfortunately have high false positive rates. An extra step is required to further analyse and filter the potential side effects detected by existing methods due to their high false positive rates, and this reduces their efficiency. In this thesis a novel methodology, the supervised adverse drug reaction predictor (SAP) framework, is presented that learns from known side effects, and identifies patterns that can be utilised to detect unknown side effects. The Bradford-Hill causality considerations are used to derive suitable attributes as inputs into a learning algorithm. Both supervised and semi-supervised techniques are investigated due to the limited number of definitively known side effects. The results showed that the SAP framework implementing a random forest classifier outperformed the existing methods on The Health Improvement Network longitudinal observational database, with AUCs ranging between 0.812-0.937, an overall MAP of 0.667, precision values between 0.733-1 and a false positive rate ≤ 0.013. When applied to the standard reference the SAP framework implementing a support vector machine obtained a MAP score of 0.490, an average AUC of 0.703 and a false positive rate of 0.16. The false positive rate is lower than that obtained by existing methods on the standard reference.
Item Type
ethesis
Thesis Type
PhD
Supervisors
Subjects (LC)
Associated Schools / Departments
School of Computer Science (UK)
eprints ID
14360
UoN Repository URI
Except where otherwise noted, this item's license is described as
File(s)![Thumbnail Image]()
Name
thesis.pdf
Type
Full-text
Size
5.43 MB
Format
Adobe PDF
Checksum ()