(149) Development of Algorithms for the Implementation and Detection of Adverse Events in Routine Electronic Data Sources

Background: Pharmacoepidemiological studies using electronic data bases require valid algorithms to detect phenotypes, such as adverse events (AEs), as clinical endpoints. In the German health care setting, new opportunities are arising to conduct such studies, with the federally funded Medical Informatics Initiative (MII) supporting infrastructural development.

Objectives: To develop and establish content validity of a set of algorithms to detect AEs using one or more data categories for implementation in electronic medical records (EMR) or statutory health insurance claims data.

Methods: The selection of AEs for algorithm development was based on a previously conducted RAND survey, in which 42 potentially drug-related AEs were prioritized by expert consensus according to their overall importance for medication safety. Six further AEs were additionally considered relevant by our research group. The two-stage algorithm development process involved multidisciplinary expert teams (clinical pharmacists and pharmacologists, medical specialists and laboratory physicians) from 10 university sites collaborating within the Use Case POLAR of the MII. For each AE, a lead centre created an initial algorithm that was either based on already existing algorithms (identified through structured literature research), or (if not available) was newly developed. A second centre was responsible for review and revision. The algorithms were finalised by group consensus. The main data categories used for the development of algorithms were ICD-10 codes for diagnoses, ATC codes for medication (for the treatment of AEs), LOINC for laboratory values, OPS codes for medical and surgical procedures, and medical findings (e.g., blood pressure).

Results: A total of 82 algorithms were developed for 48 AEs, i.e. an average of two per AE, which differ either in the type or number of data categories used (56 single-category and 26 more complex algorithms consisting of ≥ 2 categories) or in the code selection within the same data category. ICD-10 codes are the most frequently used data category, required for 65 AE algorithms. LOINC are part of 22, ATC codes of 18, OPS codes of 11, and medical findings of 6 algorithms. All algorithms were modelled in the ontology-based TOP framework which enables the execution of phenotypic queries, and will be made publicly available through a GitHub repository.

Conclusions: In collaboration with numerous experts in Germany, we have developed algorithms for a large number of AEs, which can be implemented in routine electronic data sources and serve as a basis for further medication safety research. As a next step, we will empirically validate the algorithms against all information (including free text) contained in electronic medical records.