Alarm fatigue in pediatric hospital wards is driven by frequent alarms that do not result in clinical action. Nonactionable alarms increase workload and can desensitize staff to important alerts. Identifying which alarms are likely to trigger clinical responses is necessary to optimize monitoring strategies and reduce the overall alarm burden.
The study aimed to identify factors associated with alarms that lead to clinical interventions and to develop a predictive model to support an action-oriented approach to monitoring reduction in pediatric wards.
This investigation was a prospective cohort study conducted in two pediatric wards of a tertiary hospital during 2018–2019. The researchers created a person-day level dataset by linking three routinely collected sources: bedside alarm logs, records of alarm-triggered responses, and patients' background clinical information.
An alarm-triggered intervention was defined as either physician paging or a therapeutic action taken following an alarm event. The analytic approach used a generalized linear mixed-effects logistic regression model to derive a risk score for the probability that a person-day would include an alarm-triggered intervention. Model performance was evaluated using 10-fold cross-validation.
The dataset encompassed 1,049 person-days from 286 patients. Alarm-triggered interventions occurred on 299 person-days, representing 28.5% of person-days.
The final predictive model incorporated both patient-level current-day variables and previous-day alarm/response metrics. Overall model discrimination was reported as adequate, with an area under the receiver operating characteristic curve (AUC) of 0.82. Calibration was described as acceptable based on the cross-validation assessment.
A simulation using the model to identify and discontinue monitoring on low-risk days resulted in a 25.6% reduction in alarms. In that simulated scenario, no unplanned pediatric intensive care unit (PICU) transfers were omitted, as reported in the study abstract.
The final model selected the following current-day patient predictors:
In addition, the model used prior-day alarm and response indicators:
These combined predictors were used to estimate the daily risk that an alarm would trigger a clinical intervention.
Model discrimination on cross-validation was quantified with an AUC of 0.82, indicating adequate ability to distinguish person-days with versus without alarm-triggered interventions. Calibration was reported as acceptable. The model was derived and internally evaluated using 10-fold cross-validation; external validation was not reported in the provided abstract.
To evaluate potential operational impact, the investigators simulated discontinuing physiologic monitoring on days the model classified as low risk. This simulation reduced the total number of alarms by 25.6% and, according to the abstract, did not result in omission of any unplanned PICU transfers in the simulated dataset. The abstract does not provide additional simulation parameters, thresholds used to define low-risk days, or details on safety monitoring during discontinuation.
Using routinely available bedside alarm logs, response records, and patient background data, the authors identified factors associated with alarms that trigger physician paging or therapeutic actions. The resulting predictive model demonstrated adequate discrimination (AUC 0.82) and, in simulation, could reduce alarms substantially (25.6% reduction) when monitoring is discontinued on low-risk days without missing unplanned PICU transfers in the reported simulation.
These results suggest that an action-oriented, risk-based approach to monitoring decisions could help reduce alarm fatigue and the alarm burden in pediatric wards. The abstract indicates that such a model could inform bedside assessments regarding the indication for continuous monitoring.
Note: The provided source text is the article abstract. Details not reported in the abstract—such as full model coefficients, specific risk-score thresholds, exact cross-validation results beyond AUC, and external validation—were not available in the source material.