The Distributed ID3-based Decision Tree (DIDT) algorithm provides a basis for Distributed Privacy-preserving Clinical Decision Support Systems. Due to large number of features associated with clinical patient records and iterative nature of distributed algorithms, exchanging information related to all features is expensive. We show that auto-reduction for features can be achieved with significant improvement in communication costs. Auto-reduction was implemented in DIDT and results of experiments using Nationwide Inpatient Sample data sets for 2008 are presented.