Background: Continuous monitoring in paediatric intensive care units (PICUs) generates frequent clinical alarms, 87-97% of which are nonactionable, contributing to alarm fatigue and patient safety risks. At the Erasmus MC Sophia Children’s Hospital PICU, improved alarm management is needed. Machine learning offers a promising strategy to distinguish actionable from nonactionable alarms.

Objectives: The primary objective was to develop a machine learning algorithm to classify actionable alarms from multimodal vital signs. Secondary objectives were to characterise the PICU alarm burden, capture stakeholder perspectives and evaluate model feasibility for clinical application.

Methods: Retrospective Dräger monitoring data and alarms from 2,582 PICU patients (Nov 2021 – Oct 2024) were analysed. A machine learning algorithm was developed using ART M, HR, RESP and SpO2 data from 26 patients, with alarms annotated using clinical interventions, small signal deviations and temporal associations with other parameters. Logistic regression, decision tree, random forest and XGBoost were evaluated with nested cross-validation using pre-alarm features. Performance was assessed by sensitivity, specificity, balanced accuracy, with AUROC and F1-score as complementary metrics. Semi-structured interviews with four nurses and one psychologist explored alarm experiences, impacts and reduction strategies.

Results: The best model, a decision tree, showed limited performance (sensitivity 0.36-0.49, specificity 0.51-0.70, balanced accuracy ≈0.50). No discriminative features were identified, and substantial overlap and outliers limited classification. Most alarms involved SpO2 desaturation, with unit variation and temporal patterns linked to ward activity. Interviews highlighted overstimulation and desensitisation, but also that nonactionable alarms can serve as early warnings, with interpretation requiring clinical context.

Discussion: This study provides an evaluation of machine learning-based alarm classification in the PICU and integrates clinician perspectives to guide future interventions. The machine learning model was not suitable for clinical use because of retrospective labelling, the small annotated dataset and the absence of clinical context. Future research should focus on prospective annotation, larger and more diverse datasets and complementary strategies. Alarm dashboards and daily reviews are recommended to reduce alarm burden and mitigate alarm fatigue while maintaining patient safety.