Bioinformatics Methods for Learning Radiation-Induced Lung Inflammation from Heterogeneous Retrospective and Prospective Data
Figure 9
RP with a premodeling variable selection using (a), (b) the recursive feature elimination (RFE) method. Variables were chosen from a pool of 58 dosimetric, positional, and clinical variables. The top 23 variables selected by SVM-RFE are shown after applying a pruning step to correct for multicollinearity . The top 6 variables (by applying a cutoff of 5% weighting score) were used for modeling pneumonitis. (b) An SVM-RBF classifier was tested on LOO data. (c), (d) Multimetric logistic regression approach. (c) The frequency of selected models order of 3 using our two-step resampling methods. The best-selected model consisted of three parameters (D35, COM-SI, and maximum dose). (d) The results of applying the SVM methodology with RBF kernels using these selected variables on LOO testing data. Note the improved performance in this case compared to RFE variable selection.
