The implementation of electronic health records (EHRs) in German university hospitals enables the analysis of large volumes of structured routine healthcare data. The analysis of such data offers new potential for detecting and predicting adverse drug events (ADEs). The Medical Informatics Initiative (MII) plays a pivotal role in making EHR data usable for research purposes in Germany, as this initiative facilitates interoperability and exchange of EHR data from all German university hospitals. As part of the MII, the use case POLARMI (POLypharmacy, drug interActions and Risks) was developed to retrospectively identify medication-related risks in hospitalised adult patients. Within this context, the aim of this work was to develop and evaluate various methods for the detection and prediction of ADEs using routine healthcare data from several German university hospitals involved in the POLARMI project. To ensure data protection and legal compliance, parts of this work applied a distributed analysis approach, the feasibility of which was also aimed to be assessed. A set of 34 clinically relevant inpatient ADEs was identified through an expert consensus process based on the RAND®/UCLA Appropriateness Method. Key factors contributing to their overall importance were the seriousness of the ADE and its likelihood of being drug-related. In a second expert consensus process, the likelihood of specific drugs contributing to the 14 most clinically relevant inpatient ADEs was assessed in order to provide a consensus-based list of drug-event pairs indicating adverse drug reactions (ADRs). Of the 255 drug-event pairs evaluated, 2 were considered as indicators of certain ADRs, 42 as indicators of probable ADRs, 137 as indicators of possible ADRs and 74 as indicators of unlikely ADRs. Indicators of possible, probable, or certain ADRs can serve as electronic triggers for detecting ADRs in routine healthcare data, offering a basis for future research and the implementation of automated ADR detection systems. Moreover, this work provides a methodological framework for extending the developed indicator set to other ADRs through standardised consensus processes. The feasibility of developing models for ADE detection and prediction using a distributed and privacy-preserving analysis approach built upon MII interoperability standards was assessed using two ADEs as examples: gastrointestinal bleeding and drug-related hypoglycaemia. Despite several challenges such as missing laboratory data, unreliable timestamps, and heterogeneous IT infrastructures, plausible estimates for the prevalence of ADEs and regression modelling odds ratios were received. This suggests that predictive models can be developed in a distributed setting if the research question is adapted to the infrastructure and available data. Furthermore, this work demonstrated the potential of analyses within the MII. In conclusion, the expert consensus processes that resulted in content-validated ADR indicators and the evaluation of the feasibility of building models for ADE detection and prediction using a distributed analysis approach should be regarded as complementary approaches. Together, they contribute to the advancement of ADE detection and prediction, particularly within the MII. The integration of the developed ADR indicators as outcome measures in a distributed analysis approach is a future prospect, with the potential to promote drug safety studies under real-world hospital conditions.
Anna Maria Wermund (Mon,) studied this question.