Worse is better? Performance and bias implications of feature selection in radiomics-based survival analysis

We present an extension of the minimum redundancy, maximum relevance (mRMR) feature selection algorithm to accommodate right-censored survival outcomes. We use a set of 197 preoperative CT scans from patients that underwent hepatic resection to treat colorectal liver metastases, with the intention of predicting overall survival based on radiomic analysis of the largest metastasis. Using selection of the K-best features as a baseline, mRMR is introduced as a correction of the ranking, balancing relevance with the correlation of each candidate feature with those previously selected. Harrell’s C-index is introduced as a replacement for the F-statistic to accommodate right-censoring. Two pre-processing operations are considered: first, features of low univariate significance are removed; and second, features with low reproducibility in an independent data set are removed. Each algorithm was tested in a repeated 10-fold cross-validation of a Cox proportional hazards model of overall survival. We found that use of mRMR was associated with a decrease in performance C-index of -0.016. While mRMR models had lower bias, univariate significance thresholding increased the bias of mRMR models by 0.02, producing the models with highest bias. More stable feature selection methods were associated with reduced bias and better overall performance. The baseline K-best method, with no accounting for redundancy or censoring, and no pre-processing, obtained the best performance (C-index=0.600 (0.591–0.607)) and a bias of 0.000 (-0.001– 0.004).