Representativeness of a German AI-enabled data network for secondary epidemiological analysis based on electronic health records

Abstract

Introduction The ongoing digitalization of medicine, increased computing power and low-cost storage capacities enable the use of AI-based algorithms for epidemiological big data analysis of electronic patient records. The aim of this study was to evaluate the representativeness of a data network with infrastructure for federated machine learning (ML) across numerous German hospitals. This was done by comparing basic data from the ML data network with publicly available data from the Federal Statistical Office (DESTATIS) to test the scientific validity for future epidemiological analyses. Methods In a retrospective epidemiological secondary analysis, 8,106,105 case files from the ML network were examined and compared to DESTATIS data regarding age, gender, length of hospital stay, ICD-10 diagnoses, and OPS codes. In addition, ICD-10 codes for substance abuse and the regional distribution were compared to examine socioeconomic confounders. Results The variables age, gender and length of stay, as well as the most common general ICD-10 and OPS codes and ENT-specific OPS codes, showed a high level of concordance based on clinical relevance. For the ENT-specific ICD-10 codes, 2 out of 11 of the most frequent codes showed a maximum deviation of 3.71%. The analysis of socioeconomic factors and regional distribution showed no evidence for deviations. Discussion The high level of agreement for the variables examined indicates the representativeness of the ML dataset in comparison to the DESTATIS data. This finding paves the way for future epidemiological studies based on big data, which were previously unavailable in research.