Challenges in detecting and predicting adverse drug events via distributed analysis of electronic health record data from German university hospitals.

The Medical Informatics Initiative Germany (MII) aims to facilitate the interoperability and exchange of electronic health record data from all German university hospitals. The MII use case "POLyphamacy, drug interActions and Risks" (POLAR_MI) was designed to retrospectively detect medication-related risks in adult inpatients. As part of POLAR_MI, we aimed to build predictive models for specific adverse events. Here, using the two adverse events gastrointestinal bleeding and drug-related hypoglycaemia as examples, we present our initial investigation to determine whether these adverse events and their associations with potential risk factors can be detected. We applied a two-step distributed analysis approach to electronic health record data from 2018 to 2021. This approach consisted of a local statistical data analysis at ten participating centres, followed by a mixed-effects meta-analysis. For each adverse event, two multivariable logistic regression models were constructed: (1) including only demographics, diagnoses and medications, and (2) extended by laboratory values. As numerically stable estimations of both models were not possible at each centre, we constructed different centre subgroups for meta-analyses. We received prevalence estimates of around 1.2% for GI bleeding and around 3.0% for drug-related hypoglycaemia. Although unavailability of laboratory values was a common reason hindering model estimation, multivariable regression models were obtained for both adverse events from several centres. Regarding our original intention to build predictive models, the median area under the receiver operating characteristic curve was above 0.70 for all multivariable regression models, indicating feasibility. In conclusion, plausible estimates for prevalence and regression modelling odds ratios were received when using a distributed analysis approach on inpatient treatment data from diverse German university hospitals. Our results suggest that the development of predictive models in a distributed setting is possible if the research question is adapted to the infrastructure and the available data.