The selective de-identification of ECGs
Date
2022
Authors
Akhtar, Musamma, author
Simske, Steven, advisor
Wang, Zhijie, committee member
Vans, Marie, committee member
Journal Title
Journal ISSN
Volume Title
Abstract
Biometrics are often used for immigration control, business applications, civil identity, and healthcare. Biometrics can also be used for authentication, monitoring (e.g., subtle changes in biometrics may have health implications), and personalized medical concerns. Increased use of biometrics creates identity vulnerability through the exposure of personal identifiable information (PII). Hence an increasing need to not only validate but secure a patient's biometric data and identity. The latter is achieved by anonymization, or de-identification, of the PII. Using Python in collaboration with the PTB-XL ECG database from Physionet, the goal of this thesis is to create "selective de-identification." When dealing with data and de-identification, clusters, or groupings, of data with similarity of content and location in feature space are created. Classes are groupings of data with content matching that of a class definition within a given tolerance and are assigned metadata. Clusters start without derived information, i.e., metadata, that is created by intelligent algorithms, and are thus considered unstructured. Clusters are then assigned to pre-defined classes based on the features they exhibit. The goal is to focus on features that identify pathology without compromising PII. Methods to classify different pathologies are explored, and the effect on PII classification is measured. The classification scheme with the highest "gain," or (improvement in pathology classification)/ (improvement in PII classification), is deemed the preferred approach. Importantly, the process outlined can be used in many other systems involving patient recordings and diagnostic-relevant data collection.
Description
Rights Access
Subject
biometrics
de-identification
pathologies
cyber security
anonymization
ECG