Ultra-high dimensional confounder selection algorithm comparison, with application to radiomics data

Abstract

Radiomics is an emerging area of medical imaging data analysis for diseases such as cancer. It involves the conversion of digital medical images into mineable ultra-high dimensional data, where, the number of extracted radiomics features is much larger than the number of observations. Machine learning algorithms are widely used in radiomics data analysis to develop powerful decision support models to improve precision in diagnosis, assessment of prognosis and prediction of therapy response. However, machine learning algorithms for causal inference have not been previously employed in radiomics analysis. In this paper, we evaluate the value of machine learning algorithms for causal inference in radiomics. We compare three recent competitive variable selection algorithms for causal inference: outcome-adaptive lasso (OAL), generalized outcome-adaptive lasso (GOAL), and causal ball screening (CBS). We used a ball covariance screening step to extend GOAL and OAL to ultra-high dimensional data, called SIS+GOAL and SIS+OAL. The proposed SIS+GOAL and SIS+OAL procedures can perform variable selection for ultra-high dimensional data analysis. Asymptotic properties of the proposed SIS+GOAL and SIS+OAL procedures are derived. We compared SIS+GOAL, SIS+OAL and CBS using simulation study and two radiomics datasets in cancer, osteosarcoma and gliosarcoma. The two radiomics studies and the simulation study identified SIS+GOAL as the optimal variable selection algorithm in the scenarios considered.