Application of complex radiation risk models and exposure uncertainty propagation to large-scale radiation cohort survival analysis with Colossus

Abstract

Studies on the effects of low-dose-rate chronic radiation exposure in populations require large amounts of data to find statistically significant associations between low radiation doses and different carcinogenic and cognitive health outcomes of interest. These studies are important not only for furthering the understanding of safe limits on Earth but also for improving the safety of missions to space. Existing statistical software packages were not designed to run complex risk models on millions of rows of data, and generally cannot consider exposure uncertainty or competing events in the data.

This need was met with the development and testing of the R package Colossus, which can apply risk models that were not typically available in R packages and take advantage of the available hardware to run regressions on big datasets quickly. One important aspect of Colossus is that it is open-source and uses a modular design to allow future capabilities to be added as needed. A series of tests were performed to verify that Colossus performed as intended and could perform analyses similar to other existing software.

The first novel application of Colossus was the analysis of large-scale plasmode simulations. A hypothetical study was proposed to characterize the competing risk effects of lung cancer mortality on the risk model parameter estimates of smoking status on Parkinson's disease mortality. Standard methods of plasmode simulations could not increase the observation time of individuals, which was hypothesized to lead to competing risk effects in the model parameter estimates. A random forest algorithm was used to extend the observation of individuals, and event rate models were simulated in the extended and unextended datasets. Analysis of the unextended data with only Parkinson's disease mortality produced results similar to the extended data with multiple competing events modeled, demonstrating that the unextended plasmode simulations induced effects similar to competing risks. Thus, simulation studies investigating the effects of competing risks in existing data require a method of extending the observation of individuals to simulate outcomes without the effects of events in the original data.

The availability of exposure uncertainty propagation methods is one of the most significant capabilities of Colossus. Functions were implemented to perform Frequentist Model Averaging and Monte Carlo Maximum Likelihood algorithms using multiple realizations of radiation exposure. These automated functions could analyze thousands of realizations several times faster than manual methods because of the large amount of time required to process the data for each realization. These capabilities allow researchers to propagate exposure uncertainties into their analyses.

Colossus can currently perform radiation cohort analysis, which was previously not available with public software; however, further work is required. Additional capabilities are planned to be added, particularly different logistic model regression options. A multiple-realization analysis with real data is planned, which will provide more meaningful regression results and a better understanding of the true uncertainty in the risk-model parameters. Finally, further work is required to optimize Colossus. Further research into available libraries and alternative code bases may open up optimizations that would allow these analyses to be performed with even less time and memory required.