Measuring the impact of anonymization on real-world consolidated health datasets engineered for secondary research use: Experiments in the context of MODELHealth project

Stavros Pitoglou; Arianna Filntisi; Athanasios Anastasiou; George K Matsopoulos; Dimitrios Koutsouris

doi:10.3389/fdgth.2022.841853

Measuring the impact of anonymization on real-world consolidated health datasets engineered for secondary research use: Experiments in the context of MODELHealth project

Front Digit Health. 2022 Sep 1:4:841853. doi: 10.3389/fdgth.2022.841853. eCollection 2022.

Authors

Stavros Pitoglou^{1

2}, Arianna Filntisi¹, Athanasios Anastasiou², George K Matsopoulos², Dimitrios Koutsouris²

Affiliations

¹ Computer Solutions SA, Research & Development Dpt., Athens, Greece.
² School of Electrical and Computer Engineering, National Technical University of Athens, Athens, Greece.

Abstract

Introduction: Electronic Health Records (EHRs) are essential data structures, enabling the sharing of valuable medical care information for a diverse patient population and being reused as input to predictive models for clinical research. However, issues such as the heterogeneity of EHR data and the potential compromisation of patient privacy inhibit the secondary use of EHR data in clinical research.

Objectives: This study aims to present the main elements of the MODELHealth project implementation and the evaluation method that was followed to assess the efficiency of its mechanism.

Methods: The MODELHealth project was implemented as an Extract-Transform-Load system that collects data from the hospital databases, performs harmonization to the HL7 FHIR standard and anonymization using the k-anonymity method, before loading the transformed data to a central repository. The integrity of the anonymization process was validated by developing a database query tool. The information loss occurring due to the anonymization was estimated with the metrics of generalized information loss, discernibility and average equivalence class size for various values of k.

Results: The average values of generalized information loss, discernibility and average equivalence class size obtained across all tested datasets and k values were 0.008473 ± 0.006216252886, 115,145,464.3 ± 79,724,196.11 and 12.1346 ± 6.76096647, correspondingly. The values of those metrics appear correlated with factors such as the k value and the dataset characteristics, as expected.

Conclusion: The experimental results of the study demonstrate that it is feasible to perform effective harmonization and anonymization on EHR data while preserving essential patient information.

Keywords: anonymization; electronic health records; harmonization; information loss; real data.