Alert Burden as a Marker of Mortality in Intermediate and Intensive Care Units

Introduction: Patients admitted to high-acuity units generate many monitoring alerts, yet temporal alert patterns of these patients remain underexplored. We sought to identify distinct alert burden patterns among intensive (ICU) and intermediate (IMCU) care unit patients using unsupervised machine learning, explore how these patterns evolve over time, and assess their potential use as predictors of mortality.


Methods: Monitoring alert data were extracted for patients admitted to three ICUs (medical, trauma, neuro-trauma) and one IMCU of an urban, level-one trauma center from April 2018 to December 2023. Data were preprocessed and grouped into sequential six-hour intervals, with each interval capturing the total number of alerts and alert rate per hour per patient. Patients with ≥5 intervals (i.e., ≥24 hours of data) were analyzed. An exponential decay model was fit to each patient’s alert data to characterize trends in alert burden over time. Estimates above the 95th percentile were excluded to reduce the influence of outliers. K-means clustering was then performed and constrained to produce clusters with a minimum of 1,000 patients for adequate group representation. Mortality differences across clusters were compared using chi-square test.. Cluster trajectories were visualized using the mean alert rate per hour across intervals.


Results: Our sample contained 78,323,184 alerts generated by 16,215 patients (mean age 6018 years, 43% female, 34% with injury-related diagnosis, 13% mortality). K-means identified two distinct alert burden patterns. Cluster A (n = 13,166) initially showed a high alert burden that declined over time, while Cluster B (n = 3,049) had a persistently elevated alert rate with minimal decay and experienced a significantly higher rate of mortality (19% vs 11%, 2(1)= 123.28, p < 0.001; Odds Ratio = 1.81; 95% CI: 1.63–2.01). Divergence of clusters was observable within 18 hours of admission.


Conclusions: Unsupervised clustering of alert data revealed two distinct alert burden patterns. Patients with persistently higher alert burden were at increased odds of mortality, indicating that alert data can be utilized soon after unit admission as a predictor of clinical outcome in high-acuity settings.