Machine learning approach for anonymizing electronic medical records

Abstract

Electronic Medical Records (EMRs) enable the sharing of patient medical data whenever it is needed and also are used as a tool for building new medical technology and patient recommendation systems. Since EMRs include patients’ private data there exist restriction to researchers for access. Thus an anonymizing technique is necessary that keeps patients’ private data safe while undamaging useful medical information. Conventional research has been focusing on de-identification which can lead to unexpected privacy exposure issue. To prevent unexpected privacy exposure issues anonymization techniques based on k-anonymity has been previously introduced. k-member clustering anonymization is a technique that approaches the k-anonymization as a clustering issue. The objective of the k-member clustering problem is to gather (i.e. cluster) records that will minimize the data distortion during data generalization process. Most of the clustering techniques include random seed selection and iteration process to gather record that gives minimum information distortion. However, dealing with massive medical patient dataset, randomly selecting a cluster seed will provide inconsistent performance. This paper proposes a seed selection method based on closeness centrality which not only provides consistent information loss but at the same time reduces the information loss and execution time. We experimentally compare our algorithm with two previous studies. The experiments show that our algorithm provides better performance with respect to information loss.