The data to information engine: Use of information systems and data to support risk assessment and outcomes research

Risk adjustment is essential for examining the effects of treatments and comparing the outcomes of care across and even within populations. Risk adjustment as carried out using administrative data bases relies on the interactive and additive associations between comorbidities and measured outcomes, such as, length of stay, readmissions, cost of care. This approach is quite unsatisfactory for giving any insights into the relationships between the care process and the outcomes measured, and explains no more than a third of the associated outcome. A better approach is the systematic capture of any explanatory data throughout the care process for use in risk adjustment. That is an approach that is used in acute care to stratify the sickest patients, such as, using APACHE II or III, or injury severity score (ISS). The approach may use clinical data, such as, vitals signs, and may be laboratory inclusive. The use of the laboratory is especially appealing because it is readily available by linkage of the laboratory to clinical data bases. The most common laboratory tests included are: pCO(2), bicarbonate, potassium, albumin, hematocrit, and white cell count. A study by Motes et al. showed that the classification of patients by severity of illness could even be done as well as by a clinical method using seven laboratory tests identified by recursive partitioning. The methods for examining the relationship of laboratory data to patient outcomes can be described in three main approaches: linear and generalized regression models for continuous variables; logistic regression for dichotomous variables; and log-linear models including a graphical ordinal model using maximal likelihood estimation for polytomous variables. Outside of these is a method of forming a continuous learning matrix which can be combined with the last method. The problem with the first method is that much of the data required for risk adjustment is not continuous. The problem with the second method is that in its most available farm it is not easily applied to a polytomous response, and it does not give an adjacent odds-ratio. I describe the application of the third method and its output to the assessment of laboratory tests and patient outcomes with three examples: fetal lung maturity and respiratory distress syndrome (RDS); and CA(125) halflife and relapse or remission of ovarian cancer with chemotherapy; relationship of nutritional risk factors to hospital length of stay. In the first case the scaled values of a surfactant test on amniotic fluid are related risk of RDS. In the second case the serum halflife of CA(125) is related to the likelihood of complete remission (but is related to the operative findings). In the third case the serum albumin, cholesterol, and clinical variables are important information for identifying risk of malnutrition.