Date

2014

Author

Öztaner, Serdar Murat

Metadata

Show full item record

Item Usage Stats

200
views

53
downloads

Cite This

Recent studies have shown that the incorporation of genomics information into the drug dosing prediction formulations increases the accuracy of the drug dosing while decreasing the frequency of adverse drug effects. The current clinical approaches for drug dosing which are supported by the best pharmacogenomics algorithms explain only some percentage of the variance in dosing. The main objective of this study is to enhance the accuracy and efficacy of the warfarin dosing algorithms by using advanced methods of data mining and estimation. A novel framework based on Bayesian Structural Equation Modeling (SEM) is proposed for warfarin dosing. The proposed framework performs better than the state-of-the-art methods which make use of linear regression such Maximum Likelihood Estimation (MLE). The Bayesian SEM is a robust and effective approach for the estimation of warfarin dosing since it facilitates the exploration and identification of hidden relationships and provides the flexibility to utilize useful prior information for achieving better prediction results. Two independent data sets are used for comparison and validation purposes in this study: The combined multi-ethnic data set provided by the International Warfarin Pharmacogenetics Consortium (IWPC) and the Turkish data set. A series of data pre-processing techniques (feature selection, data imputation) are applied on both of the v data sets which contain common set of non-genetic features and genetic features including CYP2C9 and VKORC1 as the main pharmacogenomics variables. The non-linear model has converged with coefficients having small Monte-Carlo error and absolute values consistent with prior domain knowledge. The obtained pharmacogenomics warfarin dosing algorithm based on the non-linear Bayesian Structural Equation model accounts for up to 56.7% of the variation in warfarin dosage while the referenced pharmacogenomics warfarin dosing algorithms based on the linear regression model explains up to 51.2% of the variance. The prediction performances are also improved for both the data sets (47.4% and 51.7% respectively) compared to MLE (45.1% and 49.3%).

Subject Keywords

Diabetes., Genomes., Warfarin., Pharmacogenomics., Drugs, Medical informatics.

URI

http://etd.lib.metu.edu.tr/upload/12617502/index.pdf
https://hdl.handle.net/11511/23653

Collections

Graduate School of Informatics, Thesis