Development of MPM and MPM II

The development of MPM (Lemeshowetal 1985) was distinctly different from that of both APACHE and SAPS. The first difference was that MPM was entirely statistically derived. A large number of variables (137 at admission and 75 at 24 h) were collected on 755 consecutive general medical and surgical ICU patients. Coronary care, cardiac surgery, burn, and pediatric (under 14 years) patients were excluded. Various statistical techniques were used to determine the relative importance and weight of each variable. This process allowed the developers to retain only those variables that were commonly collected, non-ambiguous, and shown statistically to have a strong association with survival status. Finally, a multiple logistic regression model was developed. Using multiple logistic regression, the developers further reduced the number of significant variables and objectively derived weights for each of those remaining. The final multiple logistic regression models directly computed an estimate of the probability of the patient dying during the hospital stay rather than a point score. The results were very compact and simple models with only seven variables at admission and seven variables at 24 h, and very little time was required to collect and record the necessary information.

A second major difference was that, while APACHE II and SAPS were performed 24 h after admission to the ICU and used the 'worst' value during the first day for each of its variables, the MPM system contained models that could be performed both immediately upon admission and at 24 h. Third, the variables were more condition based rather than the physiologically based variables predominating in APACHE II and SAPS. Fourth, the variables collected were generally 'yes' or 'no' answers.

Subsequently, MPM II (Lemeshow, etal 1993) was developed with data collected in two separate studies, again using consecutive general medical and surgical ICU

admissions excluding cardiac surgery, coronary care, burn, and pediatric (under 18 years) patients. In addition, admissions other than the first were excluded for those with multiple admissions. One of these datasets consisted of approximately 6000 cases collected at six United States hospitals. For the second dataset, the developers of MPM and SAPS (LeGall etaL 1993) joined forces to collect data on over 14 000 patients in 137 hospitals in Europe and North America during 1990

and 1991. Together, the 19 124 cases ultimately included came from a diverse sample of ICUs consisting of 41 per cent university hospitals, 27 per cent university-affiliated hospitals, 16 per cent community teaching hospitals, and 16 per cent community non-teaching hospitals. The data consisted of a large set of variables including all those used in APACHE II, the original MPM, and SAPS, together with others.

For the admission model (MPM0), the 19 124 patients were randomly separated into two groups: 12 610 cases (65 per cent) were used for model development, and 6514 (35 per cent) were used later for validation purposes. The association of each possible independent variable with hospital mortality was assessed. This list was further analyzed for frequency of inclusion, ease of interpretation, and strength of association with mortality. Through multiple logistic regression techniques, variables were further reduced and those remaining were each assigned a weight. Discrimination, or the ability to assign higher probabilities to those who died than those who survived, was assessed using the receiver operating characteristic (ROC) curve ( HanJey.and.McNeil 1982). This compares all possible pairs of a surviving and a dying patient and evaluates the proportion wherein the patient with the highest probability of mortality was the one who actually died. Calibration of the model was assessed using the Hosmer-Lemeshow goodness-of-fit test which tests the degree of correspondence of actual outcome and estimated outcome over groups of patients across the entire range of probabilities. Next, variables were assessed for removal if their elimination improved calibration while not harming discrimination. The resulting model contains 15 variables and directly computes a probability of mortality during the entire hospital stay ( Table 1).

Table 1 Variables in MPM0 (admission model) with definitions and estimated coefficients


Kicking Fear And Anxiety To The Curb

Kicking Fear And Anxiety To The Curb

Kicking Fear And Anxiety To The Curb Can Have Amazing Benefits For Your Life And Success. Learn About Calming Down And Gain Power By Learning Ways To Become Peaceful And Create Amazing Results.

Get My Free Ebook

Post a comment