Skip to Main Content
Effective resource requirement forecasting is necessary to reduce the escalating cost of care by ensuring optimum utilization and availability of scarce health resources. Patient hospital length of stay (LOS) and thus resource requirements depend on many factors including covariates representing patient characteristics such as age, gender, and diagnosis. We therefore propose the use of such covariates for better hospital capacity planning. Likewise, estimation of the patient's expected destination after discharge will help in allocating scarce community resources. Also, probable discharge destination may well affect a patient's LOS in hospital. For instance, it might be required to delay the discharge of a patient so as to make appropriate care provision in the community. A number of deterministic models such as ratio-based methods have failed to address inherent variability in complex health processes. To address such complexity, various stochastic models have therefore been proposed. However, such models fail to consider inherent heterogeneity in patient behavior. Therefore, we here use a phase-type survival tree for groups of patients that are homogeneous with respect to LOS distribution, on the basis of covariates such as time of admission, gender, and disease diagnosed; these homogeneous groups of patients can then model patient flow through a care system following stochastic pathways that are characterized by the covariates. Our phase-type model is then extended by further growing the survival tree based on covariates representing outcome measures such as treatment outcome or discharge destinations. These extended phase-type survival trees are very effective in modeling interrelationship between a patient's LOS and such outcome measures and allow us to describe patient movements through an integrated care system including hospital, social, and community components. In this paper, we first propose a generalization of the Coxian phase-type distribution to a - arkov process with more than one absorbing state; we call this the multi-absorbing state phase-type distribution. We then describe how the model can be used with the extended phase-type survival tree for forecasting hospital, social, and community care resource requirements, estimating cost of care, predicting patient demography at a given time in the future, and admission scheduling. We can, thus, provide a stochastic approach to capacity planning across complex heterogeneous care systems. The approach is illustrated using a five year retrospective data of patients admitted to the stroke unit of the Belfast City Hospital.