Balancing sensitivity and specificity: investigating the effect of age delay and critical illness events on the number of threshold alarms.

In critical care settings, continuous vital sign monitoring is crucial to ensure patient safety and timely intervention. While traditional patient monitor threshold alarm systems have been life-saving, they often generate numerous non-actionable alarms, which can overwhelm caregivers and lead to ineffective patient monitoring. We still have these numerous false alarms because we have a gap in understanding the importance of age-specific threshold settings, delay, and critical illness events inclusion in understanding the specificity and sensitivity of the threshold alarms. This study investigated the effect of age-specific thresholds, delay, and critical illness events on the number of threshold alarms to balance their specificity and sensitivity. Secondary data from 772 pediatric patients was extracted from the IMPALA Project conducted in the High Dependency Unit (HDU) at Queen Elizabeth and Zomba Central Hospitals in Malawi. Threshold crossing detector algorithms and age-defining functions were used to generate alarms and impute age-specific thresholds. Z-test was used to determine differences between normal threshold alarms and age-specific threshold alarms. Threshold alarms were categorized into different delays based on their durations to identify an adaptive delay that would minimize the threshold alarms to manageable alarms. Time series analysis was leveraged to extract and compare threshold alarms around patients with and without critical illness events per hour. Additionally, we investigated the variability of threshold alarms during the hour time windows before and after each critical illness event, considering factors such as delay and age. A multi-regression model was used to determine the effects of critical illness events on the number of threshold alarms, with a significance level set at p < 0.05, indicating statistical significance. The age-specific threshold had a positive influence on the threshold alarms by reducing the total number of threshold alarms [31.14% for ECGHR, 17.54% ECGRR and 54.79% for SPO2]. There was a greater significant difference between normal and age-specific threshold alarms (p < 0.00001). A 15-s delay reduced the total number of threshold alarms by 45%. We had more threshold alarms being generated 1 h before critical illness events occurrence, and applying delay and age-specific threshold had more impact on threshold alarms 3 h after the occurrence of critical illness events [Respiratory support (Total threshold alarms (232), 15 s delay (77), 15 s and age-specific threshold (17)] and most threshold alarms 1 h before critical illness events had longer durations. Critical illness [Convulsion (p < 0.0001), Malaria treatment (p < 0.0001), Death (p = 0.053), Respiratory support (p = 0.046), and Sepsis (p = 0.051)] had positive effects on the threshold alarm. There was a drop and increase in the vital sign values during the occurrence of these critical illness events [Bronchodilator support (β =  - 0.0030), Death (β =  - 0.0374), Malaria treatment (β =  - 0.0056), and Inotropic support (β =  - 0.0063)] indicating that more threshold alarms were produced during the occurrence of these critical illness events. Age-specific threshold, delay, and critical illness events can be used to strike a balance between the sensitivity and specificity of threshold alarms. In this way, we can reduce the number of non-actionable (false alarms) alarms and increase the number of actionable alarms around critical illness events. It is necessary to look into critical illness event alarm forecasting further.