Several statistical modeling systems have been designed to estimate the severity of illness of ICU patients. Although use of these models is becoming increasingly common, their validity over time and in a variety of ICU settings has not been rigorously tested. Between 1983 and 1985 we collected data on 2644 consecutive admissions to the ICU at Baystate Medical Center. That study resulted in the development of the Mortality Probability Model system, logistic regression models that provide a probability of hospital mortality based on data collected at ICU admission and at 24 and 48 hours in the ICU. In our current grant we are developing an updated MPM system based on a new developmental data set for 4224 patients admitted to six ICUs in four hospitals between the years 1989 and 1990. After development, we will validate the new models on a second cohort of approximately 2000 patients admitted between 1990 and 1991. We will now perform secondary analyses to determine whether the new MPM system will be effective over time in a range of ICUs. This is necessary to avoid the impractical and expensive strategy of continuously assembling new data sets comprised of new or refined variables among new or expanded groups of patients in order to keep systems such as the MPM up to date. In this grant we propose to (1) explore the effect of changes in time and patient mix on the performance of original and newly developed MPM systems; (2) explore strategies, such as customizing variable coefficients in existing models, to adjust for those effects; and (3) apply these strategies to practical issues, such as the assessment and comparison of the quality of care provided in different ICUs. For each of the 1989-1991 patients we will estimate the probability of hospital mortality using the original MPM system and the new models. Comparison of the two sets of probabilities will be performed for patients within each ICU and over all ICUs. We will also simulate """"""""new"""""""" ICUs, each with different patient profiles, so that we can begin to understand the impact that such factors as patient mix may have on goodness-of-fit tests and area under ROC curves. The effect of customizing a model's coefficients, and consequently updating the probabilities generated, as a means of adapting models to better reflect mortality in a changing ICU environment will be explored, together with the question of how many observations are needed in an ICU before new coefficients can be generated.
