Home therapist network modeling

Abstract

Home healthcare has been a growing sector of the economy over the last three decades with roughly 23,000 companies now doing business in the U.S. producing over $56 billion in combined annual revenue. As a highly fragmented market, profitability of individual companies depends on effective management and efficient operations. This dissertation aims at reducing costs and improving productivity for home healthcare companies.

The first part of the research involves the development of a new formulation for the therapist routing and scheduling problem as a mixed integer program. Given the time horizon, a set of therapists and a group of geographically dispersed patients, the objective of the model is to minimize the total cost of providing service by assigning patients to therapists while satisfying a host of constraints concerning time windows, labor regulations and contractual agreements. This problem is NP-hard and proved to be beyond the capability of commercial solvers like CPLEX. To obtain good solutions quickly, three approaches have been developed that include two heuristics and a decomposition algorithm.

The first approach is a parallel GRASP that assigns patients to multiple routes in a series of rounds. During the first round, the procedure optimizes the patient distribution among the available therapists, thus trying to reach a local optimum with respect to the combined cost of the routes. Computational results show that the parallel GRASP can reduce costs by 14.54% on average for real datasets, and works efficiently on randomly generated datasets. The second approach is a sequential GRASP that constructs one route at a time. When building a route, the procedure tracks the amount of time used by the therapists each day, giving it tight control over the treatment time distribution within a route. Computational results show that the sequential GRASP provides a cost savings of 18.09% on average for the same real datasets, but gets much better solutions with significantly less CPU for the same randomly generated datasets.

The third approach is a branch and price algorithm, which is designed to find exact optima within an acceptable amount of time. By decomposing the full problem by therapist, we obtain a series of constrained shortest path problems, which, by comparison are relatively easy to solve. Computational results show that, this approach is not efficient here because: 1) convergence of Dantzig-Wolfe decomposition is not fast enough; and 2) subproblem is strongly NP-hard and cannot be solved efficiently.

The last part of this research studies a simpler case in which all patients have fixed appointment times. The model takes the form of a large-scale mixed-integer program, and has different computational complexity when different features are considered. With the piece-wise linear cost structure, the problem is strongly NP-hard and not solvable with CPLEX for instances of realistic size. Subsequently, a rolling horizon algorithm, two relaxed mixed-integer models and a branch-and-price algorithm were developed. Computational results show that, both the rolling horizon algorithm and two relaxed mixed-integer models can solve the problem efficiently; the branch-and-price algorithm, however, is not practical again because the convergence of Dantzig-Wolfe decomposition is slow even when stabilization techniques are applied.