glenm@athena.UUCP (10/24/86)
I am wondering how the "normal" ranges for blood chemistry tests are established. I assume they use mean and standard deviation values to set up a range that includes 95% of the population. Is this really the case? If so, doesn't this approach cause problems in some cases? For example, a person whose thyroid hormone levels were near the upper end of the "normal" range would probably have more energy than one with levels near the low end of the "normal" range. Glen McCluskey ..tektronix!athena!glenm
oliver@unc.UUCP (Bill Oliver) (10/27/86)
In article <688@athena.UUCP> glenm@athena.UUCP (Glen McCluskey) writes: > >I am wondering how the "normal" ranges for blood chemistry >tests are established. I assume they use mean and standard >deviation values to set up a range that includes 95% of the >population. > >Is this really the case? If so, doesn't this approach cause >problems in some cases? For example, a person whose thyroid >hormone levels were near the upper end of the "normal" range would >probably have more energy than one with levels near the low >end of the "normal" range. > > Glen McCluskey > ..tektronix!athena!glenm For the most part, physicians in general and certainly virtually all pathologists have stopped using the term "normal" for individual laboratory results because of the problems with both the concept of "normal" meaning not indicative of a disease, and "normal" meaning having a Gaussian distribution. Instead, most folk refer to a test's "reference range". In setting up a refernce range, the investigator must first find an appropriate group of individuals representing the target population. Since most applications wish to use healthy individuals, you want to use a healthy population -- students, technicians, etc. which have been evaluated as to health and appropriate demographics. If this isn't possible (or desirable), you choose a population that is known to be free of the disease state that the given test is used to evaluate -- thus you would look for absence in liver disease in the population used to set up the reference range for AST (a liver enzyme). Some institutions use all admission lab data for new patients regardless of disease state. Once you have the test data, you look at the distribution. Some tests, such as serum sodium, potassium, bicarb, and chloride do have a Gaussian distribution at most institutions, and the reference range is set at +/- 2 standard deviations. Most tests, however, are skewed, and it is necessary to use a non-parametric evaluation. For instance, glucose, AST, and LD are skewed high and total protein, albumin, and packed cell volume are skewed low. Most of the time, a percentile ranking method is used, and the central 95% of the population makes up the reference range. In some instances, when subpopulations have known differences in distributions, they are broken out and listed separately. Thus, for some tests there are pediatric versus adult reference ranges and for others there are separate reference ranges based on sex, race, chemotherapy, or whatever. Each individual institution thus puts out different reference ranges based on the population the range is based on and how the local pathologists and clinicians have decided they want the reference range displayed (in many instances it is not necessary or desirable to actually set out a subpopulation based reference range if the clinicians are comfortable with their knowledge that a given population tends to run a little high or low). In addition, each analytic method has a different reference range. Thus, an AST run at a hospital which uses an Ektachem analyzer will have a different reference range than one which uses, say, an RA 1000 centrifugal method. Not only will the reference ranges be different, but virtually all statistical parameters will be different. This can sometimes cause problems in small hopitals which use one machine during the day when the high volume batch processing of samples is possible, and then switches to a different machine at night to cut down on overhead. It is sometimes impossible to compare values run on the different machines with each other. Deciding whether or not a single set of laboratory values indicate a disease state thus involves a bit more than simply looking to see if a value is in a given reference range. If we use a test which has, for instance, a true Gaussian distribution, we then are accepting that 5% of people without evidence of disease will be out of the reference range. In addition, since people who do have disease will have their own distribution, and this distribution will overlap with the reference range, some of the patients with the disease will fall into the reference range. Accordingly, tests are designed with particular sensitivities and specificities. Screening tests will have a rather narrow reference range, so that they will pick up most or all people with the disease (a high sensitivity), but will also call a bunch of folk without the disease out of range ( a low specificity). Those who are picked up on the screening test will then be further evaluated with a more specific (ususally more expensive) method. Again, each test method will have it's own sets of values for false positives (people without disease that are out of reference range), and false negatives (people with disease who are in reference range), and thus each test method will have it's own sensitivity, selectivity, and positive predictive value (the probabiltiy that an out of range test result will be associated with a given clinical disease state) or negative predictive value (the probability that an in range result will be associated with not having the disease). Bill Oliver, MD