Identifying Adverse Drug Reactions from Free-Text Dutch EHR in Hospitalized Patients with the Development of an Algorithm (IADRESS)

Introduction: While electronic health record (EHR) is a potentially valuable resource of adverse drug reactions (ADRs) [1,2], these ADRs are frequently not registered, registered in the wrong place or only registered using free-text entry [3,4]. Free text data cannot be managed and analyzed with mainstream software tools, but this is possible with text mining (TM) tools. Objective: To develop an algorithm to identify possible ADRs in free text of Dutch hospital EHR with a TM tool. Methods: In phase I, the previous rule-based algorithm was translated to a R-algorithm and improved it with the help of previous mentioned issues. In phase II, the terms of MedDRA and SNOMED-CT were added to identify ADRs and in phase III R-scripts were used to improve the R-algorithm. Results: In phase I, the R-algorithm identified 97% (n = 174) of the EHR notes containing possible ADRs identified by the rule-based algorithm. Five ADRs were missed compared to the rule-based algorithm, because of typo's and a EHR conversion. The last PDCA cycle compared to the golden standard achieved a sensitivity of 93%, a PPV of 11% and an F-measure of 0.2. For the potentially serious ADRs a sensitivity of 95% was achieved. There were 64 more EHR notes containing possible ADRs identified, by the R-algorithm. In phase II, there were 105 ADRs identified with the R-algorithm using MedDRA, SNOMED-CT and synonyms. In phase III, the R-algorithm improved the PPV, sensitivity and F-measure. Conclusion: The developed R-algorithm identified ADRs, however further research is required to extrapolate the algorithm and to combine it with clinical decision support systems to bring the data back to the physician to increase ADR registration.