Abstract
AbstractComputable or electronic phenotypes of patient conditions are becoming more commonplace in quality improvement and clinical research. During phenotyping algorithm validation, standard classification performance measures (i.e., sensitivity, specificity, positive predictive value, negative predictive value, and accuracy) are commonly employed. When validation is performed on a randomly sampled patient population, direct estimates of these measures are valid. However, it is common that studies will sample patients conditional on the algorithm result, leading to a form of bias known as verification bias. The presence of verification bias requires adjustment of performance measure estimates to account for this sampling bias. Herein, we describe the appropriate formulae for valid estimates of sensitivity, specificity, and accuracy to account for verification bias. We additionally present an online tool to adjust algorithm performance measures for verification bias by directly taking the sampling strategy into consideration and recommend use of this tool to properly estimate algorithm performance for phenotyping validation studies.
Publisher
Cold Spring Harbor Laboratory