Prediction or Prevention? Nurse Interactions with an Electronic Early Warning System for Fall Risk.

Background: Predictive models using machine learning technology are increasingly being incorporated into electronic health records to support staff in risk assessment and prediction of adverse outcomes. There is little research available related to how this technology fits into the nursing workflow or its effects on nurse behaviors or actual patient outcomes.

Methods: Retrospective data from four medical/surgical units were examined to explore nurse interactions with the interruptive alerts produced by the model and their chronological relation to actual falls.

Results: During the study period, 1.5% of all admissions resulted in at least one fall, and 87.0% of admissions resulted in at least one fall alert being produced by the system. Most alerts (57.3%) were dismissed by the receiver using the Snooze to Review option, and 22.0% of alerts were shown to staff members other than the primary nurse caring for the patient. Most falls (89.3%) were preceded by an alert being shown to any staff member, but a smaller number of falls (38.7%) were preceded by an alert being shown to the primary nurse.

Conclusion: In most fall cases in this sample, the primary nurse caring for the patient had never been exposed to an alert. However, most alerts were dismissed by nurses using the Snooze to Review option. Further research is needed to understand the relationship between nurse exposure to interruptive alerts and associated actions taken by nursing staff to prevent falls. Machine learning technology should be carefully studied and optimized to suit the needs and workflow of the staff and patients it is intended to serve.