Automatic Detecting Documents Containing Personal Health Information

Abstract

With the increasing usage of computers and Internet, personal health information (PHI) is distributed across multiple institutes and often scattered on multiple devices and stored in diverse formats. Non-traditional medical records such as emails and e-documents containing PHI are in a high risk of privacy leakage. We are facing the challenges of locating and managing PHI in the distributed environment. The goal of this study is to classify electronic documents into PHI and non-PHI. A supervised machine learning method was used for this text categorization task. Three classifiers: SVM, decision tree and Naive Bayesian were used and tested on three data sets. Lexical, semantic and syntactic features and their combinations were compared in terms of their effectiveness of classifying PHI documents. The results show that combining semantic and/or syntactic with lexical features is more effective than lexical features alone for PHI classification. The supervised machine learning method is effective in classifying documents into PHI and non-PHI.