Development and Validation of a Prediction Model for Alarm-Triggered Interventions in Pediatrics

Alarm fatigue in pediatric wards stems from frequent, nonactionable alarms that do not lead to clinical intervention. Predicting alarm-triggered clinical interventions is essential for optimizing monitoring strategies.

To identify factors associated with alarm-triggered interventions and to develop a predictive model to reduce the alarm burden through an action-oriented approach.

This prospective cohort study was conducted in 2 pediatric wards of a tertiary hospital (2018-2019). A person-day data set was created by linking alarm logs, alarm-triggered response records, and patients' background information. Alarm-triggered intervention was defined as either physician paging or therapeutic action following an alarm. A generalized linear mixed-effects logistic model was used to derive a risk score that was evaluated via 10-fold cross-validation.

Interventions occurred on 299 of 1049 person-days (28.5%) in 286 patients. The final model included the following predictors: aged under 1 year, underlying respiratory or neurological disease, acute respiratory tract infection, and current-day supplemental oxygen use. In addition, the model incorporated previous-day metrics, including alarm-triggered interventions or observations (bedside assessments), fewer crisis alarms (eg, arrhythmia waveforms), and frequent warning alarms (vital sign deviations). The model demonstrated adequate discrimination (area under the receiver operating characteristic curve: 0.82) with acceptable calibration. A simulated discontinuation of monitoring on low-risk days reduced alarms by 25.6% without omitting any unplanned pediatric intensive care unit transfers.

Factors associated with alarm-triggered interventions were identified using routinely available data. These findings could inform bedside assessment of monitoring indications to reduce the alarm burden.