Data-driven models for complex medical systems

This dissertation is a collection of three related but self-contained papers utilizing statistical learning methods to address operational issues in complex medical systems:; Evidence-based incentive systems with an application in health care delivery. We develop an empirical method to estimate the parameters of a multi-task principal-agent model. The principal observes the system's aggregate performance (downstream outcome) and several other performance measures (typically process-compliance measures to be referred to as intermediate outcomes). All observed measures are noisy signals of the agent's effort in each task. The principal rewards the agent based on a weighted combination of the observed performance measures. The question is to determine the optimal mix of performance measures that would maximize the principal's expected payoff. Using Empirical Likelihood, we show how the principal can use data from multiple agents to answer the following questions: (a) How can process-compliance measures be integrated into a single intermediate performance score that can be used in an optimal payment system? (b) What is the agent's cost of effort and reservation utility? (c) What is the optimal payment system? The method was applied to data from patients with kidney failure who needed dialysis (Medicare, the payer, was the principal and the dialysis providers were the agents). An optimal payment system was designed. The system was shown to have the potential to increase the number of hospital-free days per patient year-at-risk by 7.4% without increasing total medical expenses.; Understanding the relationship between dialysis facility characteristics and quality-of-care. Hospital expenditures constitute one-third of total Medicare payments for the End-Stage Renal Disease program. The existing fee-for-service reimbursement system for dialysis indirectly rewards dialysis providers for a reduction in hospital admissions through increased revenue. The Medicare Modernization Act mandated the development of new payment systems that would strengthen these incentives through pay for performance initiatives in which dialysis providers will be rewarded for reducing patient hospital admission rates. However, it is unclear whether there is an association between dialysis providers and hospitalization rates. Simple economic arguments suggest that if such an association exists it would lead to fewer hospital days for patients in non-profit dialysis facilities. Employing observational data techniques, we analyzed clinical and claims data from 170,209 Medicare eligible patients receiving dialysis in 2003 to examine the association among patient outcomes, provider characteristics and market concentration. Chief amongst our findings is that patients who dialyzed at for-profit facilities spent 21% more time in the hospital than their non-profit counterparts. In other words, 1,900 patient years in hospital and {dollar}600--900 million of inpatient costs could be averted each year if for-profit facilities were to match the hospital utilization patterns of non-profit facilities.; Optimal capacity overbooking in healthcare facilities for patients with chronic conditions. Patients suffering from a chronic condition often require periodic treatment. For example, patients with End-Stage Renal Disease (ESRD) require dialysis three times a week. These patients are also frequently hospitalized for complications from their treatment, resulting in idle capacity at the clinic. These temporary patient absences make overbooking at the clinic attractive. We develop a semi-closed migration network to capture patient flow into the clinic and between the clinic and hospital. We consider a simple class of stationary control policies for patient admissions and provide algorithms for selecting one that maximizes long-run average earnings. Local diffusion approximations were constructed to provide square-root loading formulas for the optimal capacity level and patient overb...