The use of clinical algorithms derived by linear regression analyses is a standard method in medicine, but the empirical and descriptive nature of this kind of model also results in a number of limitations and drawbacks, exemplified by the COAG and EU-PACT trials (ClinicalTrial.gov identifier NCT00839657 and NCT01119300, respectively). Regression-based models are data-driven and, therefore, population-dependent. Accordingly, they are expected to be only valid in the same patient population on which they were originally derived (i.e., mostly in Caucasians).
Indeed, such data-driven methods are at increasing risk of overfitting the data, giving rise to a so-called “prediction” model that is not generalizable to other datasets or populations. In addition, statistical analysis used to this purpose can be biased by linkage disequilibrium (LD) structure, admixture patterns and differences in minor allele frequencies (MAF) between and within populations. Population specificity is the most likely explanation to the poor predictability found in African-Americans in the COAG trial. It is becoming increasingly clear that many of the currently available pharmacogenetic-guided algorithms for warfarin dosing predictions do not work well in people with substantial African heritage.
Citation: Duconge J (2014) Population Heterogeneity and Genomic Admixture: Relevance for Global Pharmacogenetics. J Pharmacogenomics Pharmacoproteomics 5:e141. doi: 10.4172/2153-0645.1000e141