gms | German Medical Science

Conducting clinical trials requires recruiting patients in a limited time. Otherwise, evidence of the safety and effectiveness of new interventions cannot be provided. If a single hospital cannot recruit enough patients, multicenter trials are means of best choice, and potential partners must be identified. The Medical Informatics Initiative (MII) in Germany is establishing data integration centers to exchange healthcare-related data and the German research data portal for health [1]. Based on the idea of this portal, we developed APERITIF [2] for case number estimation. Using the trial’s free-text eligibility criteria, a query is generated and sent to a server to identify the number of potential trial subjects in a hospital’s database.

Our approach is based on the MII core dataset containing modules such as Diagnosis or Procedure [3]. The corresponding data elements are specified according to the standard Fast Healthcare Interoperability Resources (FHIR) [4]. We used Clinical Quality Language (CQL) [5] as query language as it supports FHIR, and it is more expressive for eligibility criteria than FHIR Search.

Medical and demographic data are extracted from eligibility criteria using Natural Language Processing methods and the Unified Medical Language System based entity recognition MetaMapLite including negation detection [6]. For medical concepts, SNOMED CT and LOINC codes [4] are requested at an instance of the Ontoserver, a terminology server based on FHIR [7]. We also used Expression Constraint Language (ECL) for requesting subconcepts in SNOMED CT [8]. The corresponding CQL query is formulated and sent to an instance of Blaze, a FHIR server able to process CQL queries [9].

We analyzed 552 eligibility criteria of 27 studies to identify commonly used concepts: disorders, therapies, medications, laboratory results, age, and gender (about 84 % of all concepts). For the implementation, we considered the corresponding codes and data elements of the core dataset of the MII and developed our program APERITIF.

Negations, entities, and junctions of 20 CQL queries were rated by four study nurses of the campus Lübeck, Germany. A precision of 62.64 % (inclusion criteria) and 66.45 % (exclusion criteria) was achieved and mainly reduced by errors in entity recognition.

As no test data was available, we manually generated patients and medical data for the 20 studies to check whether junctions and negations work correctly. At least one patient per study should be identified by the query. For ten queries, the corresponding test patient could be retrieved successfully.

Inclusion criteria are often more complex compared to exclusion criteria. Therefore, entity recognition is more difficult for these criteria and requires further modifications to improve the precision.

Four queries containing too many laboratory codes caused parsing exceptions by the Blaze. Hence, the number of codes per laboratory entity was restricted, although information gets lost. Existing contradictions for five queries could be resolved in favor of inclusion criteria. Only one contradiction remained after these modifications because of an error in negation detection.

We successfully demonstrated an approach to perform feasibility queries for subject estimation using innovative methods such as CQL, Blaze, Ontoserver, and ECL.

The authors declare that they have no competing interests.

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

This contribution has already been published [10].