It is hoped that anesthesiologists and other clinicians will be able to increasingly rely upon laboratory test data to improve the perioperative care of patients. However, it has been suggested that in order for a laboratory test to have clinically useful diagnostic performance characteristics (sensitivity and specificity), its performance must be considerably better than those that have been evaluated in most etiologic or epidemiologic studies. This pessimism about the clinical utility of laboratory tests is based upon the untested assumption that laboratory data are normally distributed within case and control populations.We evaluated the data distribution for 700 commonly ordered laboratory tests, and found that the vast majority (99%) do not have a normal distribution. The deviation from normal was most pronounced at extreme values, which had a large quantitative effect on laboratory test performance. At the sensitivity and specificity values required for diagnostic utility, the minimum required odds ratios for laboratory tests with a nonnormal data distribution were significantly smaller (by orders of magnitude) than for tests with a normal distribution.By evaluating the effect that the data distribution has on laboratory test performance, we have arrived at the more optimistic outlook that it is feasible to produce laboratory tests with diagnostically useful performance characteristics. We also show that moderate errors in the classification of outcome variables (e.g., death vs. survival at a specified end point) have a small impact on test performance, which is of importance for outcomes research that uses anesthesia information management systems. Because these analyses typically seek to identify factors associated with an undesirable outcome, the data distributions of the independent variables need to be considered when interpreting the odds ratios obtained from such investigations.
ASJC Scopus subject areas
- Anesthesiology and Pain Medicine