Evaluation of Pediatric eCART Implementation

Pediatric eCART draws upon readily available EHR data and rapidly quantifies disease severity, predicting the likelihood of critical illness onset. Currently, no consistently available system continuously tracks the risk of critical illness in children admitted to UW Health. While AFCH has an implementation of Pediatric Early Warning Scores (PEWS) available for risk monitoring, internal reports indicate limited usage. Therefore, AFCH/UW Health clinicians or care providers do not have a reliable mechanism to risk-stratify patients for effective clinical decision-making.

This proposal leverages the AgileMD clinical decision support engine and a machine learning analytic developed in a dataset of over 30,000 patients. Pediatric eCART was explicitly designed to draw attention to patients at increased risk of deterioration and optimize patient management, including the timing of and need for ICU-level care.

Preliminary studies indicate that pediatric eCART implementation at the University of Chicago has led to improved outcomes. Similar improvements among children admitted to UW Health will lead to decreased morbidity and mortality among the pediatric population.

Further, a significant gap in understanding of nurse acceptance of data-driven CDS tools remains. Nurses are the largest workforce of clinicians in the health system and play a primary role in the detection of clinical deterioration as the clinicians that spend the most time observing and assessing patients; however, AI-driven CDS acceptability has not been measured to assess nurse acceptance of these emerging tools. Acceptability is essential to increase sustained use and to decrease suboptimal outcomes such as alert fatigue or increased cognitive load so that these tools ultimately mediate nurse well-being. One study assessed nurse perceptions of the usefulness of a sepsis early warning system and found that less than half of nurses perceived the alerts to be helpful and only a third of nurses reported that the alerts impacted patient care. Understanding nurse acceptance will inform AgileMD's design strategies to foster uptake and use so that predictive tools may be leveraged to improve the cognitive burden of nurse clinicians. In the end, the study will evaluate pediatric eCART on two pediatric groups: (1) screened pediatric patients; (2) pediatric nurse clinician end-users.

Study Design: This is a pre- and post- interventional study of a machine learning algorithm integrated into the electronic health record as a clinical decision support tool. The "pre" participants are hospitalized children (less than 18 years old) who were admitted to UW Health between January 1, 2022, and the date of pediatric eCART implementation in 2025. Pediatric eCART scores will be retrospectively calculated for the "pre" participants by feeding a patient's labs and vital sign observation into the pediatric eCART tool. The "post" participants are hospitalized children (less than18 years old) who will be admitted to UW Health within the two years following pediatric eCART implementation (expected 2025-2027). Pediatric eCART scores will be calculated in real-time for these patients.

• Median 30-day Ventilator-Free Days [Time frame: assessed through hospital stay (typically up to 5 days on average, but may be over 60 days)]
• Summary of Critical Events [Time frame: assessed through hospital stay (typically up to 5 days on average, but may be over 60 days)]
• Total Hospital Length of Stay (LOS) [Time frame: assessed through hospital stay (typically up to 5 days on average, but may be over 60 days)]
• Number of ICU transfers [Time frame: assessed through hospital stay (typically up to 5 days on average, but may be over 60 days)]
• Usability of Pediatric eCART: System Usability Scale (SUS) score [Time frame: Surveys automatically sent to nurses within a week of eCART interface, responses collected up to 1 month]
• Acceptability of Pediatric eCART: Perceived Usefulness Scale [Time frame: Surveys automatically sent to nurses within a week of eCART interface, responses collected up to 1 month]