ESSAYS IN HEALTHCARE MODELING AND ANALYTICS

Restricted (Penn State Only)
Author:
Agrawal, Deepak
Graduate Program:
Industrial Engineering
Degree:
Doctor of Philosophy
Document Type:
Dissertation
Date of Defense:
May 26, 2017
Committee Members:
  • Soundar Kumara, Dissertation Advisor
  • Soundar Kumara, Committee Chair
  • Vinayak V Shanbhag, Committee Member
  • Guodong Pang, Committee Member
  • Mort D Webster, Outside Member
  • Guodong Pang, Committee Chair
  • Guodong Pang, Dissertation Advisor
Keywords:
  • appointment scheduling
  • healthcare analytics
  • preference based scheduling
  • markov decision process
  • no-show
  • multi-priority
  • network model
  • risk prediction
  • physician scheduling
Abstract:
In spite of increasing costs the quality of care in the US is decreasing therefore, the President’s Council of Advisors on Science and Technology (PCAST) report on healthcare in 2014 emphasized the need for applying systems engineering principles and analytics to healthcare operations improvement. This along with the availability of increasing amount of healthcare data has created a renewed interest in researchers to develop efficient and practical healthcare decision support systems. To that end, in this dissertation we study three important, related, and yet independent, healthcare research problems are studied with two viewpoints data analytics based and operations research (OR) based techniques in healthcare. In this research, our objective is to develop, validate and integrate the data-driven predictive models into the hospital workflow and provide decision making support for medical professionals. It is our hope this research can help generate solutions to these problems, improving the efficiency of hospital operations, patient outcomes, and patient and staff satisfaction. The three issues addressed in this dissertation are: (1) patients’ return to hospital within 72 hours after discharge from being an inpatient (e.g., surgical) due to adverse events, (2) developing a provider scheduling system for outpatient clinic at a multi-clinic facility, and (3) developing a preference based scheduling model for a healthcare provider network (HPN). In the first problem we develop a stacked classification model to predict short-term and long-term readmission, emergency department revisits and deaths (REDD) along with identifying the important factors contributing to adverse events. We proposed a stacked classification model with level one training models using logistic regression, random forest and gradient boosting method. The second level learning model was a non-negative least square method. The final outcome is a linear combination of the predictions from the three learning models. The short-term (3-day) model had a fair c-statistic of 0.671 and 0.664 for the derivation and validation cohort, respectively. The c-statistic of long-term (30-day) REDD model was 0.713 and 0.711 for the derivation and validation cohort, respectively. These models were implemented at Geisinger Health System and are under pilot study. In the second problem, we propose a multi-objective provider scheduling model under demand and capacity uncertainty and patient no-shows. Though the problem is formulated as a mixed integer programming (MIP), the network structure of the problem simplifies the computational complexity of the problem. Resource flexibility is an important criterion for any scheduling model to be practical. In the final research problem, the appointment scheduling with multiple clinic locations, multiple physicians is proposed as a network model, which gives it the necessary flexibility to dynamically assign providers to patients in real time but also to be rigid enough so that it can be implemented months in advance. In addition to that, we consider multi-priority patients with heterogeneous preference in our model. We develop theoretical bounds for the best achievable policy as well propose a heuristic approach to solve the dynamic problem. Using extensive numerical experiments we show the proposed scheduling policies can outperform the current scheduling practices. Finally, to make the model closer to reality we propose two extensions to the model. In the first extension, the demand follows doubly stochastic Poisson process. We also show that the proposed dynamic policy is robust under the assumption of a doubly stochastic assumption. In the second extension, we propose joint capacity planning and appointment scheduling problem, which can be formulated as a mixed-integer nonlinear programming (MINLP). iv