Workload, Predictive Accuracy, and the Value of Algorithm-Enabled Process Innovation : The Case of Sepsis

Predictive algorithms have an increasingly important role in supporting the day-to-day operations of healthcare organizations. Yet, fully realizing the value of algorithms lies critically in the opportunity to re-engineer the related processes and redefine roles in ways that make organizations more effective. We study how and when algorithm-enabled process innovation (AEPI) creates value in light of dynamic operational environments (i.e., workload) and behavioral responses to algorithmic predictions (i.e., algorithmic accuracy). Our context is an AEPI effort around a rule-based algorithm for early detection of sepsis|an infection in the bloodstream. Sepsis is a costly condition that is the leading cause of death for hospitalized patients. It is estimated that 11.0 million people died from sepsis globally in 2017. We collaborated with a large U.S. based hospital system and examined whether AEPI developed for sepsis care (sepsis AEPI) impacts patient mortality and when this impact is stronger or weaker. We utilize a rich set of clinical and nonclinical data in empirically examining the impact of sepsis AEPI on patient mortality. We leverage the staggered implementation of sepsis AEPI across hospital units and conduct our estimation on a carefully matched sample. The matching utilizes data on patient vitals and the logic behind the algorithm to create a robust comparison group consisting of patient visits for which sepsis AEPI would have triggered an alert if it had been in place. Our empirical analysis shows that sepsis AEPI reduces the likelihood of death from sepsis (43% relative reduction in mortality risk due to sepsis). A higher than usual workload and an increase in the average number of inaccurate alert experience at a hospital unit (e.g., an oncology unit, which provides care for cancer patients), in general, reduces the effectiveness of AEPI. We also identify diminishing mortality benefits over prolonged periods of adoption; evaluation of the moderators over time helps explain this diminishing impact. Our findings suggest that streamlining sepsis care processes through a predictive algorithm (i.e., algorithm-based monitoring of real-time patient data and providing predictions, streamlined communication channels for coordinating care for a patient with sepsis prediction, and a more standardized process for sepsis diagnosis and treatment) holds significant promise in reducing the loss of life from sepsis. We show that such value, however, is sensitive to operational and behavioral factors as the algorithm becomes a routine part of the day-to-day operations of the hospital