Introduction: The PRECISE trial uses a validated machine learning (ML) algorithm to identify a subgroup of infected patients who have a potential mortality benefit from balanced crystalloids based on their initial vital signs. When a clinician orders normal saline (NS) on a patient from this subgroup, a clinical decision support system (CDSS) recommends changing the order to balanced crystalloids. The aim of this study is to evaluate whether the non-knowledge-based (ML) CDSS could impact a clinician’s future fluid ordering behavior. Methods: From June 13 to September 30, 2024, patients were automatically enrolled in the PRECISE trial if they were at Emory Healthcare with a suspected infection, received a NS order, and were identified by a ML algorithm as potentially benefiting from balanced crystalloids. The change in the percentage of NS orders was evaluated by comparing the three months before and after the alert interaction. A chi-squared test was used to evaluate at the department level and a paired t-test to evaluate at the provider level. The analysis included all crystalloid orders, regardless of whether the patient had an infection. Results: There were 12,735 NS orders placed before clinicians interacted with the alert and 12,169 NS orders placed after interaction with the alert. Most of the orders were placed in the ED (9,071 before, 8,875 after) compared to the ICU (138 before, 161 after). When evaluating at the collective group level, the overall percentage of NS bolus and NS continuous intravenous (IV) fluid orders after alert interaction did not change ordering practices (ICU + ED p=0.95, ICU p=0.73, ED p=0.15). The top 10% of clinicians placing NS orders, “NS superusers” did have an overall decrease in the percentage of NS orders after alert interaction by 3.4% (p< 0.0001). At the clinician level, there was a decrease in percentage of NS orders after alert interaction in the ICU (p=0.036) and “NS superusers” (p=0.039), but no difference was seen overall (p=0.29) or in the ED (p=0.28,). Conclusions: “NS superusers” had a higher rate of initial interaction with the alert which may prompt them to modify their ordering behavior to avoid future alerts. The low volume of NS orders from the ICU group makes this finding less impactful and requires further investigation.
Coopersmith et al. (Sun,) studied this question.