Dynamic distributed predictive learning models that preserve privacy for hospitals with insufficient labeled data

A prediction model built dynamically using patient data from multiple hospitals can serve as a tool for suggestive knowledge in clinical decision support. Such a tool that accommodates queries based on attributes of interest is helpful in building a targeted model from multiple hospitals when a local clinical data repository does not have sufficient number of records to draw conclusions from. However, because of privacy concerns and legal ramifications, hospitals are reluctant to divulge raw medical records. Hence, mechanisms to build distributed prediction models using just the statistics of patient data are attractive. Distributed ID3-based decision tree (DIDT) algorithm is such a prediction model builder. In this study, we analyze National Inpatient Sample data for 3 years and demonstrate that DIDT can be used to help collaboratively build better predictive models when hospitals have insufficient number of records for good local models. Using 261 attributes for model building, we showed that collaborating hospitals with less than 100 cases of hospitalizations for a targeted disease were able to achieve good improvement in accuracies for predicting hospitalization collectively using a distributed model compared to local models. When relying on local models for predicting risks for sample diseases, more patients were misclassified and some local patients could not be classified. Our collaborative model effectively reduced misclassification providing accurate early diagnostics to additional patients. The profile of hospitals with sufficiently large number of patient records was explored to identify local models with specific characteristics that can serve the needs of hospitals with insufficient data.