gms | German Medical Science

Introduction: Manually creating realistic software tests for complex data is time consuming [1]. At a minimum, cases to be covered correspond to the number of execution paths that depend on the input data [2], [3]. A formal specification or source code for automatically generating tests may not be available, as in the case of test-driven development or proprietary software [4]. Further, for cross-institutional collaboration on clinical data and software, patient privacy prohibits sharing real cases [5]. Hence, while the benefit from testing for correctness and development rises with the realism of test data [6], [7], practitioners resort to synthetic data in this domain.

Since medication (the administration of drugs) can greatly change the state of a patient, its record is an essential part of clinical data. Medication data can also be fairly complex; it may include a mixture of drugs, each consisting of several ingredients, varying quantity administered over time, and additional information such as dosage form and application site. This structure combines hierarchical information and corresponds to complex, multivariate probability distributions.

Methods: We capture medication data using deep generative models [8], [9]. This allows the creation of new, synthetic data from the learned distribution, which ideally matches the original data distribution without disclosing any personally identifiable information [10]. Its value further increases with an accurate representation of relevant features. Therefore, we evaluate statistical properties of simulated data and compare them to that of original medications. In this way, and in addition to the use case of software testing, we can also assess the possibility of sharing complex, simulated medication data for pooled analytics that add to single-site studies while preserving patient privacy.

Conclusion: In summary, we generate complex, simulated medication data with software testing as a first use case, and evaluate its suitability for future use in pooled, privacy-preserving analytics.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.