gms | German Medical Science

Introduction: The IT-based detection of adverse events has great potential benefits for patient safety. The detection of adverse events, defined as any untoward medical occurrence in a patient who has received a drug, requires the consideration of several related signs of different types. In particular, the co-occurrence of abnormal clinical parameters and medication use is of great interest, as this can detect drug-related problems (DRP). DRP cause preventable negative health outcomes for patients [1], [2]. So far, gathering data for detecting adverse events is carried out manually by reviewing and collecting data from clinical narrative notes and electronic records. This is typically difficult, time-consuming, and costly [3], [4]. Therefore, it is necessary to apply automated techniques to analyze EHRs and identify a given set of observations that form an adverse event of interest. In many cases, this means searching for patterns of triggers that suggest a DRP. Our aim is to clarify how adverse events can be algorithmically identified using MII-CDS specified FHIR resources.

Method: First, we made assumptions from the healthcare process on dependent and independent observations that might qualify as triggers. Second, we modeled the trigger-generating process and researched all information sources. Third, we mapped all information sources to adequate FHIR resources in conjunction with the Medical Informatics Initiative Core Data Set (MII CDS). Fourth, we listed all observations and their conditions that make them considerable as triggers. Fifth, the relation between the observations, i.e., the concurrency of the observations, was specified to form a trigger. Seventh, we mapped the triggers and their constraints to query able FHIR^® resources. Lastly, we mapped each set of observations and their conditions for a specific healthcare setting of interest to the timestamps of the FHIR resources within defined timespans from the intervals. The intervals represent the physiological processes mapped onto the observations documented in the healthcare process. The patterns found can then be presented to healthcare professionals for validation as drug-related problems.

Results: We found that there is no easy way to construct an algorithm to pass to FHIR-based data. Hence, we had to select FHIR resources for our problem, i.e., MedicationRequest with Medication, Condition, Observation and Procedure and identify for each observable and for each scenario the necessary timestamps. Then we created a set of rules per resource, per observable and per scenario that detect if one or more triggers are active and might lead to drug-related problems. This is carried out using an expert consensus process survey.

Discussion: Gathering insights from healthcare services is not a simple task, as assumptions about the physiological and service processes need to be represented by the documented data. Our approach to detecting adverse events in EMR data requires knowledge about the process and decisions inflicting potential adverse events and their timing in a real-world setting. For more complex adverse events, it might be necessary to combine the detection algorithm of single observations with ontological representations, as done using the TOP-Framework [5].

The authors declare that they have no competing interests.

The authors declare that a positive ethics committee vote has been obtained.