Innovative Fall Prevention: UCHealth’s New AI Approach
Tackling Patient Falls in Healthcare Settings
Inpatient clinical care teams have long utilized various tools—such as bed alarms, gait belts, and strategically placed nursing stations—to combat the pressing issue of patient falls. However, despite these interventions, many healthcare professionals face challenges such as workflow disruptions and alert fatigue. This diminishing effectiveness has led to a heightened need for innovative solutions in fall prevention strategies.
Advancing Technology for Better Care
Responding to this need, UCHealth, based in Aurora, Colorado, has developed a groundbreaking user interface. This novel tool tailors specific fall interventions based on a patient’s individualized risk profile, ensuring that only those at a high risk receive intensive precautions. This not only aims to enhance patient safety but also seeks to alleviate alarm fatigue experienced by clinicians.
A Data-Driven Solution
The UCHealth system employs a unique tool that adopts a data-centric approach. It utilizes mobility data, behavioral health indicators, and other essential risk factors to forecast the likelihood of inpatient falls, particularly those that could result in injury. Integrating these predictions into the Epic electronic health records (EHR) system facilitates a seamless workflow for clinical teams.
Extracting Meaningful Insights from EHRs
As Brendan Drew, a data scientist at UCHealth, states, the goal is to distill the significant amount of data inputted into the EHR systems into actionable and clear recommendations. This effort is supported by Brittany Cyriacks, a clinical informaticist at the health system, who underlines the importance of creating clear communication channels for clinicians regarding these predictions.
The Journey Towards Effective Model Creation
A multidisciplinary team undertook an extensive literature review to develop the predictive model for fall risk. They successfully identified 12 risk domains and 92 potential variables, ultimately sampling data from over 181,000 inpatient admissions. Each admission was assessed using more than 200 features, which were instrumental in understanding fall risks associated with various patient profiles.
Modeling Methods: From Complexity to Simplicity
The initial model tests employed two methodologies: XGBoost and a regularized logistic regression model. After several iterations, the top features were streamlined and incorporated into a more simplified logistic regression model. This model updates every four hours, incorporating the most recent clinical documentation to provide timely risk assessments.
Comprehensive Risk Classifications
Rather than relying solely on binary high/low indicators, UCHealth’s tool offers a three-tiered classification system—Highest Risk, Elevated Risk, and Universal Risk. This stratified approach provides healthcare teams with more clinically relevant results and recommendations for patient safety measures.
Challenges in Clinical Workflow Integration
While the model shows promise, integrating this predictive tool into existing clinical workflows has not been without challenges. The team needed to establish clear precautions correlating with each risk level, ensuring that staff members were equipped with specific actions based on the classification of each patient.
Key Challenges Addressed
Meaningful Risk Display: Simply showcasing a numeric risk score was inadequate—clinicians required insights into why a patient was flagged as high risk along with recommended precautions.
Initial Data Availability: The first 12 hours of a patient’s admission can present limited data. Therefore, staff must apply clinical judgment to determine whether to implement universal precautions while awaiting comprehensive categorization from the model.
Fostering User Trust: For the model to be effective, acceptance and trust among nursing and care team members are vital. Ongoing training and a user-friendly design have been pivotal in enhancing confidence in the AI system.
- Seamless Workflow Integration: To prevent adding extra steps for clinicians, UCHealth has embedded the model’s outputs and recommendations directly into existing Epic flowsheets, pop-ups, and care plans. This integration allows for a natural utilization of the tool during patient care.
Enlightening Future AI Applications
At the upcoming HIMSS25 conference in Las Vegas, Drew and Cyriacks will delve deeper into their experiences with artificial intelligence in fall prevention. The session titled "Fall Injury Risk Model: From AI to Clinical Interventions" promises to provide valuable insights and lessons learned from their endeavor.
Practical Takeaways for the Healthcare Community
During the conference session, attendees can expect to learn about strategies for effective adoption of predictive models, performance assessments, and methods for addressing healthcare disparities—for instance, ensuring that resources are inclusive for Spanish-speaking patients.
Commitment to Continuous Improvement
UCHealth’s approach emphasizes a commitment to ongoing refinement and engagement from both organizational leadership and nursing teams. Their experiences thus far focus on the collaborative efforts necessary to integrate AI tools and foster a strong safety culture among clinical staff.
Conclusion
UCHealth’s innovative approach to predicting and preventing falls in healthcare settings represents a significant leap forward in patient safety. By combining state-of-the-art artificial intelligence with a nuanced understanding of patient risk factors, the health system is setting a standard for others to follow. The insights offered at HIMSS25 will not only guide fellow healthcare professionals but will also shape the future of patient care, ensuring that safety measures serve their intended purpose efficiently and effectively.
