No abstract

No abstract

Introduction Healthcare services are complex and challenging to measure. Traditionally, healthcare delivery has been evaluated by three categories of measurement: structure, process and outcome (Donabedian, 1980). The progress report "America's best hospitals" released annually since 1990 uses these three quality dimensions to rate the best hospitals in the USA (US News & World Report, 1990). The human and material resources available in each hospital are used to assess the structure of the hospitals. Outcomes are usually evaluated by the standardized mortality ratio (SMR) which is the ratio of the observed to expected mortality rate in each hospital. However, the process dimension has been more difficult to measure; a survey conducted amongst physicians from the American Medical Association asked them to name five "best" hospitals in their respective field using process as the primary factor, but without giving any guidance for which specific measures to use or what techniques should be used to measure them; this highly subjective assessment did not reach a consensus (US News & World Report, 1996). However, the expert stakeholders in this study identified "patient care" and "patient comfort" as the most significant process measures. Until recently measuring the performance of these was done by prognostic scoring systems such as the "acute physiology and chronic health evaluation" (APACHE); the simplified acute physiology score or the "mortality prediction model" (Zimmerman, 2002). All of these systems consider binomial patient outcome namely "survival" or "death" as the indicators of measurement. These systems incorporate logistic regression equations to predict the mortality for a case-mix in a particular intensive care unit (ICU). The ratio of the predicted mortality to the observed mortality (SMR) is used to compare the performance of different ICUs (Becker and Zimmerman, 1996). Although used by many studies, there are many inherent problems with these models; a study which has used all the three models to compare ICUs from 32 hospitals (Project IMPACT) reported only a fair-to-moderate agreement in the identification of quality measures (Glance et al., 2002). Other studies have reported poor goodness of fit for these scoring systems, implying that the prognostic models do not perform consistently in all ICUs (Marik and Varon, 1999; Markgraf et al., 2000; Katsaragakis et al., 2000). Although patient outcome should always be the primary goal of any ICU, there are many other contributory factors that also have to be considered which are omitted from these scoring systems. Attempts have been made to resolve these omissions using such methods as data envelopment analysis which have helped to improve structural measures (Field and Emrouznejad, 2003; Dlugacz et al., 2002). Changes in individual ICU outcome factors such as an increase in "mortality rate" could be interpreted as a reduction in the level of overall performance. However, it is imperative that before any conclusions are finalised all possible causes must...

DEA, Efficiency, Ratio analysis, Weight restrictions, C61, C14,

Traditionally most cross-selling models in retail banking use demographics information and interactions with marketing as input to statistical models or machine learning algorithms to predict whether a customer is willing to purchase a given financial product or not. We overcome with such limitation by building several models that also use several years of account transaction data. The objective of this study is to analysis credit card transactions of customers, in order to come up with a good prediction in cross-selling products. We use deep-learning algorithm to analyze almost 800,000 credit cards transactions. The results show that such unique data contains valuable information on the customers’ consumption behavior and it can significantly increase the predictive accuracy of a cross-selling model. In summary, we develop an auto-encoder to extract features from the transaction data and use them as input to a classifier. We demonstrate that such features also have predictive power that enhances the performance of the cross-selling model even further.