Patient Flow Modeling in the Long-term Care Network

Hambisa A Keno, Purdue University

Abstract

Long-term care (LTC) represents an expensive segment of the healthcare delivery system. The range of health and social services delivered under LTC span across diverse points of care and across a much larger timescale than the typical intra-facility, single episode of care. An approach for model-based policy decision in LTC delivery is characterizing patient flow in the LTC network and evaluating policy alternatives as scenario variants of the basic flow model. A Length of stay (LOS) based modeling framework that accommodates special properties of the patient flow problem in the LTC network was proposed. These properties include censored duration observations, competing transition states, recurrent transition events, importance of past transition histories and non-Markovian sojourn time distributions under a fully parametric survival analysis model. Patient transitions among facilities in the LTC network were found non-Markovian. Gamma frailty components were found significant in characterizing recurrent transition events in the LTC network. Regression parameters for spatial-temporal history variables were also found significant. Moreover, influential factors for nursing home care utilization were identified using the proposed framework. A pattern of flow into facilities of high or low care intensities was detected which was dependent on clinical profile and care intensity at current facility. The LOS-based model was extended to characterize patient flow across the entire LTC network as a Multi-State Semi-Markov (MSSM) model. Here, the order of the embedded Markov chain that best characterized the patient transition data was investigated. To limit the set of regression parameters in the MSSM model only to causal variables, a Bayesian Network Structure Learning method was applied on each arc in the LTC network. The sensitivity of parameters describing the MSSM model to the order of the embedded Markov chain was diagnosed. The goodness-of-fit for higher-order embedded Markov chains monotonously increased up to order three. Parameters for sojourn time in care facilities of higher intensity, namely hospital and nursing home, were found sensitive to the order of the embedded Markov Chain. The MSSM model was embedded in a state-transition simulation model. Methods for characterizing the semi-Markov transition process as a two-stage model as well as translating the Cox Proportional Hazards (PH) model into a simulation tool for rescaling flow parameters for heterogeneous patients was proposed. Parameter variation experiments using the simulation model identified dementia diagnosis and dual eligibility as critical variables to increased LTC utilization. Finally, upper bounds for expected cost-savings from statewide interventions were projected using the simulation model.

Degree

Ph.D.

Advisors

Kong, Purdue University.

Subject Area

Industrial engineering|Health care management|Systems science

COinS