The selection of variables and the assignment of weights to levels of these variables in SAPS II were accomplished with the assistance of the logistic regression modeling technique. This approach differs from that used in older systems, which were based on clinical judgment alone. Collecting the data necessary to calculate the SAPS II score is very simple and rapid. We estimate that it would take less than 5 min per patient. All the variables in the SAPS II are readily available and do not require special venous or arterial blood samples.
One of the goals of the modeling process was to maintain a pure physiology-based system. However, criteria of calibration and discrimination were improved considerably by including the three underlying chronic clinical conditions. As with any system based on clinical measurements, missing values present a problem which has been identified by others. This problem is generally solved by assuming that values not recorded in the medical record are within normal limits. These rules were followed with SAPS II since, for instance, in some countries serum bilirubin is not systematically measured. Similarly, blood gases are not necessarily measured in all non-ventilated patients.
Some systems require that a single diagnosis be specified for estimating the probability of mortality. In SAPS II the probability of mortality is calculated directly from the score using a logistic regression equation, without adding points or any sort of correction for the acute disease. This decision was made before beginning the study, based on the assumption that selection of a single diagnosis is too difficult for most ICU patients. Although some patients can be categorized according to a specific, simple, and unique diagnosis, such as chronic obstructive pulmonary disease, septic shock, or barbiturate overdose, this is not generally the case. In fact, it has been found possible to categorize only 37 per cent of patients into a single diagnostic category, with the remaining patients having multiple diagnoses ( BahJouJ §1
al 1988). While knowledge of diagnosis would certainly have an impact on the estimated probability of mortality, such estimates will be available for only a small percentage of ICU patients. When several diagnoses are possible, it is often difficult to select the most important. For example, if a patient has acute respiratory distress syndrome and associated purulent peritonitis, which is the main diagnosis? Other systems require specification of the principle diagnosis, and risk of death will differ according to the chosen category. Before it can be determined that adjustment for diagnosis has been successful in estimating risk, the calibration of the model must be carefully checked within diagnosis groups.
A more complete discussion can be found in the original article ( LeGall etaL 1993).
Was this article helpful?