Proactive Coordination in Healthcare Service Systems through Near Real-Time Analytics

The United States (U.S.) healthcare system is the most expensive in the world. To improve the quality and safety of care, health information technology (HIT) is broadly adopted in hospitals. While EHR systems form a critical data backbone for the facility, we need improved 'work-flow' coordination tools and platforms that can enhance real-time situational awareness and facilitate effective management of resources for enhanced and efficient care. Especially, these IT systems are mostly applied for reactive management of care services and are lacking when they come to improving the real-time "operational intelligence" of service networks that promote efficiency and quality of operations in a proactive manner. In particular, we leverage operations research and predictive analytics techniques to develop proactive coordination mechanisms and decision methods to improve the operational efficiency of bed management service in the network spanning the emergency department (ED) to inpatient units (IUs) in a hospital, a key component of healthcare in most hospitals. The purpose of this study is to deepen our knowledge on proactive coordination empowered by predictive analytics in dynamic healthcare environments populated by clinically heterogeneous patients with individual information changing throughout ED caregiving processes. To enable proactive coordination for improved resource allocation and patient flow in the ED-IU network, we address two components of modeling/analysis tasks, i.e., the design of coordination mechanisms and the generation of future state information for ED patients. First, we explore the benefits of early task initiation for the service network spanning the emergency department (ED) and inpatient units (IUs) within a hospital. In particular, we investigate the value of proactive inpatient bed request signals from the ED to reduce ED patient boarding. Using data from a major healthcare system, we show that the EDs suffer from severe crowding and boarding not necessarily due to high IU bed occupancy but due to poor coordination of IU bed management activity. The proposed proactive IU bed allocation scheme addresses this coordination requirement without requiring additional staff resources. While the modeling framework is designed based on the inclusion of two analytical requirements, i.e., ED disposition decision prediction and remaining ED length of stay (LoS) estimation, the framework also accounts for imperfect patient disposition predictions and multiple patient sources (besides ED) to IUs. The ED-IU network setting is modeled as a fork-join queueing system. Unlike typical fork-join queue structures that respond identically to a transition, the proposed system exhibits state-dependent transition behaviors as a function of the types of entities being processed in servers. We characterize the state sets and sequences to facilitate analytical tractability. The proposed proactive bed allocation strategy can lead to significant reductions in bed allocation delay for ED patients (up to ~50%), while not increasing delays for other IU admission sources. We also demonstrate that benefits of proactive coordination can be attained even in the absence of highly accurate models for predicting ED patient dispositions. The insights from our models should give confidence to hospital managers in embracing proactive coordination and adaptive work flow technologies enabled by modern health IT systems. Second, we investigate the quantitative modeling that analyzes the patterns of decreasing uncertainty in ED patient disposition decision making throughout the course of ED caregiving processes. The classification task of ED disposition decision prediction can be evaluated as a hierarchical classification problem, while dealing with temporal evolution and buildup of clinical information throughout the ED caregiving processes. Four different time stages within the ED course (registration, triage, first lab/imaging orders, and first lab/imaging results) are identified as the main milestone care stages. The study took place at an academic urban level 1 trauma center with an annual census of 100,000. Data for the modeling was extracted from all ED visits between May 2014 and April 2016. Both a hierarchical disposition class structure and a progressive prediction modeling approach are introduced and combined to fully facilitate the operationalization of prediction results. Multinomial logistic regression models are built for carrying out the predictions under three different classification group structures: (1) discharge vs. admission, (2) discharge vs. observation unit vs. inpatient unit, and (3) discharge vs. observation unit vs. general practice unit vs. telemetry unit vs. intensive care unit. We characterize how the accumulation of clinical information for ED patients throughout the ED caregiving processes can help improve prediction results for the three-different class groups. Each class group can enable and contribute to unique proactive coordination strategies according to the obtained future state information and prediction quality, to enhance the quality of care and operational efficiency around the ED. We also reveal that for different disposition classes, the prediction quality evolution behaves in its own unique way according to the gain of relevant information. (Abstract shortened by ProQuest.