gms | German Medical Science

Introduction: In intensive care units (ICUs) patients' vital signs are constantly monitored and personnel is alarmed if the measurements exceed given thresholds. Up to 94% of the alarms require no intervention, causing stress and alarm fatigue in the staff [1]. Numerous Research projects aim to reduce this high rate of non-actionable alarms [2].

Medication management is the main action after an alarm aiming to adjust patients' vitals [3]. Medication records in the hospital information system are therefore invaluable for machine learning (ML) approaches to tackle non-actionable alarms. We aim to examine a large dataset for existing relationships between different alarm types and medication actions in their temporal proximity, or whether manual documentation error noise or other data quality issues obscure them.

Methods: During this retrospective, observational study (IRB approval EA1/127/18), we collected alarm and patient data from 19 adult ICUs of a large German hospital between 2019 and 2021 and assembled them in a database with a modified MIMIC-IV schema [4]. This analysis focuses on medication actions in the context of red alarms (highly critical) of five physical alarm conditions (PAC) [5]: low oxygen saturation (SpO₂), low or high heart frequency (HF) and low or high blood pressure (BP). Medication actions were defined as start or stop of an administration, increase or reduction of a dosage within a time window of 15 minutes after an alarm. We filtered out regular actions prior to the analysis. For each PAC, we queried medication actions and calculated the cumulated percentage (cut-off set at 70%). Medications appearing within at least one PAC medication list were included in our descriptive analysis.

Results: For 35,004 patients (40,865 stays), we counted 3,987,543 medication actions and 7,949,606 alarm start log entries. Red alarms make 9.93%, the remainder are yellow (less critical). We counted 661 different substances applied in the time window. Using the 70% cutoff, 43 substances remained across all PACs (Table 1 [Tab. 1]).

Some medication action distributions are highly skewed towards a single PAC, while others have a rather flat distribution across PACs (e.g. with norepinephrine, heparin). Overall distribution shows a distinct non-random structure.

Discussion: Some distributions are clinically explicable- e.g., amiodarone, magnesium, salbutamol, or urapidil; these medications were most likely applied due to medical indication after an alarm. The administration of substances such as norepinephrine, fluids, or propofol is often routine – which explains their more even distribution – but their variance may represent clinically meaningful responses to specific PACs. Some other substances seem to occur randomly due to routine applications. The start and increase actions show a clearer pattern, compared to the stop and decrease actions, which are more random and sometimes contradictory.

Conclusion: To our knowledge, this is the first analysis examining medication actions in the context of patient monitoring alarms. In addition to providing clinical insights, it showed that there is a sensible non-random structure in the collected data. In a next step the dataset will be used for ML to predict actionable alarms.

Grant: BMBF 16SV8559.

The authors declare that they have no competing interests.

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