gms | German Medical Science

Introduction: The AKTIN (Alliance for information and communication technology in intensive care and emergency medicine) emergency department data registry is a federated and distributed research infrastructure that gives access to electronic health records from German emergency departments (EDs) [1]. Data are stored within each participating ED node in an AKTIN data warehouse (DWH). Data can be queried using the AKTIN Broker.

Distributed research networks like the German Medical Informatics Initiative [2] or the Norwegian PraksisNett [3] rely on the efficient data access authorization of research requests. In the AKTIN registry, a technical process for federated data access authorization was established [1]. When EDs do not respond to queries, registry personnel manually contact them. While non-essential alerts may lower the responsiveness of clinical personnel [4], this consideration typically receives little attention within research infrastructures and networks.

The objective of this work was to cluster the users of the federated data access authorization process used in the AKTIN registry. As we suspect users to exhibit consistent patterns in their response time, we employed process mining to group ED nodes based on their temporal behavior [5]. Grouping nodes with similar approval times could allow for tailored notifications that could reduce bottlenecks and improve user response times.

Methods: We employed process mining [5] to cluster the users of the federated approval process based on user interactions stemming from the communication log files of the AKTIN Broker between 2017 and April 15th, 2024. After extracting log files from the AKTIN Broker, we transformed them into an event-log for process mining [5]. The AKTIN DWH allows for automatic approval of periodically repeated queries. We excluded these data and focused exclusively on queries that underwent manual review studying manual response behavior. We used an inductive miner to generate process models. We clustered users based on the median time of transition upon approval of a query from the event ‘query received’ by an AKTIN DWH to ‘query queued’ as the median provides a robust measure for typical node behavior.

Results: We observed n = 4,054 queries communicated to 56 ED nodes. The AKTIN DWH nodes can be clustered into four groups, maximizing cluster cohesion and separation as measured by the silhouette score. One group of 17 nodes completes queries on median after 10 (IQR: 3) days, one of 5 nodes after 25 (IQR: 4) days, one of 32 nodes after 2 (IQR: 4) days and one of 2 nodes after 41 (IQR: 0) days.

Discussion and conclusion: We clustered users of the federated approval workflow into four groups of nodes. Those with fast and medium interaction times answer reliably while a third group answers typically in batches. The last group consisted of two new ED nodes of the same hospital responded jointly. While we may overlook the full variability and specific outlier by focusing on median response time, the results can be used to develop tailored notifications. These align with the FAIR principles (Findability, Accessibility, Interoperability, and Reusability) by improving the manageability and usability of ED routine data.

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