In this work, we study the behavior of a feature selection algorithm (backwards selection) using random forests, by fusing multi-modal data from different subjects. Two separate datasets related to cutaneous melanoma, obtained from image (dermoscopy) and non-image (microarray) sources are used. Imputations are applied in order to acquire a unified dataset, prior the effect of machine learning algorithms. The results suggest that application of the normal random imputation method acts as an additional variation factor, helping towards stability of potential recommended biomarkers. In addition, microarray-derived features were favorably selected as best predictors compared to image-derived features.