pearcedom/survivALL: Continuous Biomarker Assessment by Exhaustive Survival Analysis

In routine practice, biomarker performance is calculated by splitting a patient cohort at some arbitrary level, often by median gene expression. The logic behind this is to divide patients into “high” or “low” expression groups that in turn correlate with either good or poor prognosis. However, this median-split approach assumes that the data set composition adheres to a strict 1:1 proportion of high vs. low expression, that for every one “low” there is an equivalent “high”. In reality, data sets are often heterogeneous in their composition (Perou, CM et al., 2000 <doi:10.1038/35021093>)- i.e. this 1:1 relationship is unlikely to exist and the true relationship unknown. Given this limitation, it remains difficult to determine where the most significant separation should be made. For example, estrogen receptor (ER) status determined by immunohistochemistry is standard practice in predicting hormone therapy response, where ER is found in an ~1:3 ratio (-:+) in the population (Selli, C et al., 2016 <doi:10.1186/s13058-016-0779-0>). We would expect therefore, upon dividing patients by ER expression, 25% to be classified “low” and 75% “high”, and an otherwise 50-50 split to incorrectly classify 25% of our patient cohort, rendering our survival estimate under powered. 'survivALL' is a data-driven approach to calculate the relative survival estimates for all possible points of separation - i.e. at all possible ratios of “high” vs. “low” - allowing a measure’s relationship with survival to be more reliably determined and quantified. We see this as a solution to a flaw in common research practice, namely the failure of a true biomarker as part of a meta-analysis.

install.packages("remotes")
remotes::install_github("pearcedom/survivALL")