Saving Lives With Analytics: Incorporating the Human-technology Frontier in the Design of Alert Systems for Early Detection of Sepsis

By addressing key issues on both sides of the human-technology frontier, this dissertation focuses on improving care quality for a prevalent, deadly, and costly condition: sepsis. Sepsis is the body’s overwhelming response to infection, affecting nearly two million people annually and accounting for half of all hospital deaths in the United States. Detecting sepsis early and providing timely treatment can significantly impact patient outcomes. To facilitate early detection, healthcare providers are increasingly leveraging automated sepsis alert tools. In close collaboration with a large hospital group in the Midwest U.S., we sought to address the shortcomings of an alert system commonly employed in hospitals while also enhancing its capabilities to improve sepsis care processes. In the first part, focusing on the human side of alert systems, we empirically study how clinical teams provide care in compliance with evidence-based standards using an alert system. Specifically, we consider clinical teams consisting of two roles, nurse and physician, and investigate the impact of nurses’ timely completion of alert-related tasks (i.e., acknowledging the alert and notifying physicians within a designated time frame) on physicians’ compliance with sepsis care standards. We find that nurses’ timely response to alerts has a positive spillover effect on physicians’ compliance, and this effect becomes stronger as workload increases and weaker as the number of false alerts increases. In the second part, we design an alert system for early detection of sepsis. Our design accounts for both the technology and human sides of alert systems, which together determine the efficacy of alerts. On the technology side, we personalize alerts based on a patient’s individual characteristics, thus improving the accuracy of alerts. On the human side, we acknowledge the role of users—caregivers—and account for their compliance behavior in following care standards embedded in alerts’ workflow. Taking into account these two aspects (i.e., accurate prediction on the technology side and timely action on the human side), we develop an optimization framework in which predictive models are integrated with a prescriptive model to determine when to give a warning to caregivers about the possible presence of sepsis. Using clinical data from our partner hospital, we validate the performance of our personalized and compliance-aware design. Our design detects more sepsis cases and alerts earlier than the hospital’s alert system, resulting in higher quality-adjusted life days for patients. In the third part, focusing on the technology side of alert systems, we use data readily available in the electronic health records to develop an intelligible machine learning algorithm that predicts sepsis. In particular, we use the explainable boosting machine classifier to predict sepsis based on patients’ individual characteristics and physiological measurements. We then show that our algorithm provides accuracy comparable to state-of-the-art (unintelligible) machine learning approaches. Alert systems can help healthcare providers deliver high-quality sepsis care only if they provide accurate and timely alerts, smoothly integrate with clinical workflows, and ensure caregivers’ compliance with embedded guidelines. By considering these key determinants of an effective alert system, this dissertation takes a step toward improving sepsis-care quality.