Abstract
Multimodal data allows supervised learning while considering multiple complementary views of a problem, improving final diagnostic performance of trained models. Data modalities that are missing or difficult to obtain in clinical situations can still be incorporated into model training using the learning using privileged information (LUPI) framework. However, noisy or redundant features in the privileged modality space can limit the amount of knowledge transferred to the diagnostic model during the LUPI learning process. We consider the problem of selecting desirable features from both standard features which are available during both model training and testing, and privileged features which are only available during model training. A novel filter feature selection method named NMIFS+ is introduced that considers redundancy between standard and privileged feature spaces. The algorithm is evaluated on two disease classification datasets with privileged modalities. Results demonstrate an improvement in diagnostic performance over comparable filter selection algorithms.