Regularized mixture cure models for improving cancer risk stratification systems by identifying latent subgroups

Music City Center 

Many researchers have sought to identify prognostic models for time-to-event outcomes. When the covariate space is high-dimensional, penalized Cox proportional hazards (PH) models are commonly fit. However, various groups have shown that advances in therapy for leukemia and myelodysplastic syndrome have increased overall survival rates and some groups have identified factors related to long-term survival, defined as survival exceeding three years, such that patients on newer treatment regimens were more likely to be long-term survivors. In fact, some argue that these improved outcomes indicate that some acute myeloid leukemia (AML) patients can be considered "potentially cured." The mixture cure model (MCM) is a time-to-event model that is used when a cured fraction exists. MCMs assume the population consists of two subgroups, those cured will not experience the event of interest and those susceptible to the event of interest. Therefore, "cured" can be considered synonymous with attaining long-term relapse-free survival. MCMs are more appropriate than the Cox PH model when the dataset includes a cured fraction, because in such situations the Cox PH model does not accurately estimate the hazard or survival primarily because the assumption of a constant hazard ratio over time is violated. Moreover, the two regression components in MCMs permit identification of features associated with cure and/or latency of susceptible patients. However, few variable selection methods exist for MCMs, especially for high-dimensional covariate spaces when there are more predictors than samples. In this talk, I will describe our recent work, namely, the development of regularized MCMs for high-dimensional covariate spaces, which allow for (1) the identification of prognostic factors associated with both cure status and/or latency as well as (2) identification of latent subgroups that may be useful in the development of cancer risk stratification systems. Our methods have been made available in our R package, hdcuremodels.