Distributed Privacy-Preserving Decision Support System for Highly Imbalanced Clinical Data

When a medical practitioner encounters a patient with rare symptoms that translates to rare occurrences in the local database, it is quite valuable to draw conclusions collectively from such occurrences in other hospitals. However, for such rare conditions, there will be a huge imbalance in classes among the relevant base population. Due to regulations and privacy concerns, collecting data from other hospitals will be problematic. Consequently, distributed decision support systems that can use just the statistics of data from multiple hospitals are valuable. We present a system that can collectively build a distributed classification model dynamically without the need of patient data from each site in the case of imbalanced data. The system uses a voting ensemble of experts for the decision model. The imbalance condition and number of experts can be determined by the system. Since only statistics of the data and no raw data are required by the system, patient privacy issues are addressed. We demonstrate the outlined principles using the Nationwide Inpatient Sample (NIS) database. Results of experiments conducted on 7,810,762 patients from 1050 hospitals show improvement of 13.68% to 24.46% in balanced prediction accuracy using our model over the baseline model, illustrating the effectiveness of the proposed methodology.