Minimizing Misclassification Cost in Healthcare Associated Infection

Motivated by healthcare associated infection (HAI), which is estimated to cost US hospitals $9.8 billion per year, we develop a machine-learning based scheme to administer patient-targeted preventive measures upon admissions to the hospital. The model involves an objective function that acknowledges the asymmetry between the cost of missing an infection and the cost of a ”false alarm,” the former (real cost to the hospital) could be order-of-magnitude higher than the latter (minor expenses for certain preventive measures, which are nonetheless imperfect). A two-step algorithm is devised to solve the problem in a much more efficient manner than a black-box algorithm such as deep neutral network, and with a clearly interpretable solution. We also provide convergence and rate-of-convergence results, in the form of distribution-free probabilistic guarantees, using a variation of the Dvoretzkey-Kiefer-Wolfowitz bound (as refined by Massart). We illustrate the performance of the model with real data from several largest hospitals in New York City.