Introduction: Electronic health record alerts in the ICU aim to help clinicians make medical decisions and improve patient safety. However, repeated alerts may be habitually ignored, reducing their effectiveness. We sought to measure how many times alerts fire in the ICU before a physician places an order in response and hypothesized there would be variability in likelihood of response across alert types. Methods: We analyzed EHR data on interruptive alerts sent to residents and attending physicians in 7 adult medical and surgical ICUs at University of Michigan from 6/2022-10/2024. Immediately actionable alerts were those in which an order can be placed within the alert. We calculated the order conversion rate (percent of alerts fired that had an order placed within the alert). Among alerts that were ultimately acted upon, we calculated the number of firings before an order was placed. We used logistic regression to estimate the likelihood that an order would be placed based on the number of previous firings without response, adjusting for the specific alert type. Results: Among 151,257 actionable alerts, 131,694 (87.1%) were sent to 1,313 residents and 19,563 (12.9%) to 1,012 attending physicians during 11,278 ICU admissions. 50 distinct actionable alert types were identified. The most common alerts recommended handwashing signs for Clostridioides difficile (12.7%), use of a pain medication order set rather than individual opioid orders (12.5%), and use of a UTI-prevention order set (9.9%). The overall order conversion rate was 8.2%. Among ultimately acted-on alerts, providers saw an alert 4 times on average per patient before placing an order (median 3, IQR 1-4). After adjusting for alert type, the predicted probability of placing an order with the first alert was 9.1%. This decreases by 0.5% with every additional alert exposure. The odds of placing an order in response to an alert varied significantly by alert type (e.g. transfusion-related alert with 67-fold higher odds of order placement vs reference standard alert). Conclusions: Overall, the first firing of an actionable alert results in orders placed less than 10% of the time and decreases with subsequent firings. Actionable alerts with low likelihood of having an order placed should be evaluated for re-design or de-implementation.
Hechtman et al. (Sun,) studied this question.