Access Type

Open Access Dissertation

Date of Award

January 2014

Degree Type


Degree Name



Industrial and Manufacturing Engineering

First Advisor

Ratna B. Chinnam

Second Advisor

Alper Murat




This dissertation research investigates opportunities for developing effective decision support models that exploit near real-time (NRT) information to enhance the "operational intelligence" within hospital Emergency Departments (ED). Approaching from a systems engineering perspective, the study proposes a novel decision support framework for streamlining ED patient flow that employs machine learning, statistical and operations research methods to facilitate its operationalization.

ED crowding has become the subject of significant public and academic attention, and it is known to cause a number of adverse outcomes to the patients, ED staff as well as hospital revenues. Despite many efforts to investigate the causes, consequences and interventions for ED overcrowding in the past two decades, scientific knowledge remains limited in regards to strategies and pragmatic approaches that actually improve patient flow in EDs.

Motivated by the gaps in research, we develop a near real-time triage decision support system to reduce ED boarding and improve ED patient flow. The proposed system is a novel variant of a newsvendor modeling framework that integrates patient admission probability prediction within a proactive ward-bed reservation system to improve the effectiveness of bed coordination efforts and reduce boarding times for ED patients along with the resulting costs. Specifically, we propose a cost-sensitive bed reservation policy that recommends optimal bed reservation times for patients right during triage. The policy relies on classifiers that estimate the probability that the ED patient will be admitted using the patient information collected and readily available at triage or right after. The policy is cost-sensitive in that it accounts for costs associated with patient admission prediction misclassification as well as costs associated with incorrectly selecting the reservation time.

To achieve the objective of this work, we also addressed two secondary objectives: first, development of models to predict the admission likelihood and target admission wards of ED patients; second, development of models to estimate length-of-stay (LOS) of ED patients. For the first secondary objective, we develop an algorithm that incorporates feature selection into a state-of-the-art and powerful probabilistic Bayesian classification method: multi-class relevance vector machine. For the second objective, we investigated the performance of hazard rate models (in particual, the non-parametric Cox proportional hazard model, parametric hazard rate models, as well as artificial neural networks for modeling the hazard rate) to estimate ED LOS by using the information that is available at triage or right after as the covariates in the models.

The proposed models are tested using extensive historical data from several U.S. Department of Veterans Affairs Medical Centers (VAMCs) in the Mid-West. The Case Study using historical data from a VAMC demonstrates that applying the proposed framework leads to significant savings associated with reduced boarding times, in particular, for smaller wards with high levels of utilization.

For theory, our primary contribution is the development of a cost sensitive ward-bed reservation model that effectively accounts for various costs and uncertainties. This work also contributes to the development of an integrated feature selection method for classification by developing and validating the mathematical derivation for feature selection during mRVM learning. Another contribution stems from investigating how much the ED LOS estimation can be improved by incorporating the information regarding ED orderable item lists.

Overall, this work is a successful application of mixed methods of operation research, machine learning and statistics to the important domain of health care system efficiency improvement.