Guest Comment

Tberchtold

Description

Guest CommentarySpectrum Bias and Clinical Decision MakingStuart Spitalnic, MDlinical decision making often includes the use ache but normal results on neurologic examinationof tests to either confirm or exclude conditions and CT scan, what is the probability of a subarachnoidfrom a differential diagnosis. Generally, if a test hemorrhage (ie, what is the posttest probability)?C of adequate specificity is positive in a patientCONDITIONAL PROBABILITYlikely to have the condition being tested for, the diagno-sis is essentially confirmed. By the same token, if a test You consult the general formula for conditionalof adequate sensitivity is negative in a patient with a low probability (Bayes’ formula), which is listed below.likelihood of having the condition being tested for, theP [B|A] × P [A]diagnosis is essentially excluded. Unfortunately, the sim-P [A|B] =plicity of the above rules is disturbed by the concept of P [B|A] × P [A]+P [B|A] × P [A]spectrum bias. Spectrum bias results from the fact that atest’s sensitivity and specificity are not fixed values but This equation states that the probability of A given Brather vary with the severity or temporal stage of the dis- equals the probability of B given A multiplied by theease being considered. Failure to take spectrum bias probability of A, divided by the sum of the probability ofinto account might result in using test results to erro- B given A multiplied by the probability of A and theneously confirm or ...

Informations

• Publié par: Ziom
• Nombre de lectures: 212
• Langue: English

Extrait

22 Hospital PhysicianNovember 2001

Dr. Spitalnic is an Assistant Residency Director, Brown University Emergency Medicine Residency Program, and an Assistant Professor of Medicine, Brown University, Providence, RI; he is also a member of theHospital Physician

G u e s tC o m m e n t a r y

e  

-

 tivity o rhage i detecti You sults sh

| | | |

ut normal results on neurologic examination scan, what is the probability of a subarachnoid rhage (ie, what is the posttest probability)?

w

|

|

s -

Stuart Spitalnic, MD

Spectrum Bias and Clinical Decision Making

S p i t a l n i c: Sp e c t r u mB i a s: pp .2 2 Ð 2 4

(1 Ðsensitivity) × (prevalence) count is reasonably expected i = (1 Ðsensitivity) × (prevalence) + (specificity) × (1 Ðprevalenceappendicitis; unfortunately,) have tion does not consider the spe .05 × .01 = .0005 = .05% Although it may be reasonable t .05 × .01 + .99 × .99 leukocyte count in a patient wit According to this calculation, if a patient has negativedicitis who is developing perit results on a CT scan, the probability of there being aexpect an elevated count in some subarachnoid hemorrhage is only 1 in 2000, which isfirst twinges of pain associated considerably lower than the pretest probability of 1 inappendix. Both patients have app 100. Incorporating this newly acquired information intoat opposite ends of the spectru your clinical practice, you decide not to perform a lum-find in the literature a specific p bar puncture on the patient, because a negative resultsensitivity of the leukocyte count on a CT scan nearly excludes a subarachnoid hemor-dicitis, you rarely find it reporte rhage from the differential diagnosis. Furthermore, youties for early and late presentatio stop performing lumbar punctures on all patients withThe concept of spectrum bias severe headache who have similarly negative results onerical values. Consider 2 patients CT scan. You later learn that several of these patientswho is hypoxic, febrile, and pro went on to develop subarachnoid hemorrhage. Review-tum and the other who is just ing your files, you discover that you evaluated 500 peo-Would you expect a chest radiogr ple for severe headache, 5 of whom eventually had asensitivity in each of these pati subarachnoid hemorrhage (or 1%, as would have beenpneumonia, but the second patie predicted by the prevalence); the diagnosis was missedobvious changes on a chest radio on all 5 of their CT scans. What went wrong?Subarachnoid hemorrhage ( the outset of this discussion) is a THE CONCEPT OF SPECTRUM BIAS from a small sentinel bleed with Be assured, especially if you took the time to com-headache) to a catastrophic blee mit BayesÕ formula to memory, that the formula andrologic deterioration, herniation the logic of the arguments above are all correct. Whatin this range represent subarac is missing from the previous determinations is a consid-and all are likely to be included eration of the possibility of spectrum bias. As previouslydetermine the sensitivity of CT sc stated, spectrum bias implies that the sensitivity andarachnoid hemorrhage. However specificity of diagnostic tests vary with the severity oris comatose after having a subar duration of a given medical condition. For example,may have a large bleed (and ther you can determine the sensitivity of measuring the cre-a positive result on a CT scan), atine kinase (myocardial bound) (CKMB) level forheadache but no abnormalities detection of myocardial infarction in a patient in youric examination may have a very emergency department by calculating the percentageand a small bleed can be below t of patients with elevated CKMB levels in whom myocar-by a CT scan. Consequently, a C dial infarction has been confirmed. Standard proce-ent sensitivities in patients at di dure for ruling out myocardial infarction, however, re-spectrum of the disease. A CT s quires the measurement of serial CKMB values overfor disease in 100% of patient time, because CKMB has different sensitivities at differ-owing to a subarachnoid hemorr ent times. Very early in a myocardial infarction, mea-sensitivity of only 10% in those surement of CKMB is an insensitive measure but, aftersive symptoms. If, instead of the several hours, becomes very sensitive. Although a sin-scans for detecting subarachnoid gle sensitivity for measurement of CKMB in myocardialat the beginning of this discussio infarction can be quoted, it is actually more accurate to10% (representing the sensitivity express this sensitivity over the spectrum of time.ing subarachnoid hemorrhage i There are many examples of how spectrum biasthe mild end of the spectrum affects the performance of a test at different stages ofwould arrive at a posttest probabil an illness. Consider the use of the leukocyte count inof subarachnoid hemorrhage giv the evaluation of appendicitis. An elevated leukocytea CT scan) close to 1%, the obser

www.turnerwhite.com

Hospital PhysicianNovember 2001 23

S p i t a l n i c: Sp e c t r u mB i a s: pp .2 2 Ð 2 4

CIRCUMVENTING SPECTRUM BIAS

How can you avoid falling into the evidence trap sur- rounding spectrum bias when you apply test perfor- mance data to your patient population? As with most problems in interpretation of medical literature, the answer can be found by carefully reviewing the methods of the sources of the sensitivity and specificity data. Before incorporating the test into your practice, make sure you know what the sensitivity and specificity of the test isin patients like the one you are testing. When, for exam- ple, you are evaluating a patient for deep vein thrombo- sis, determine if the sensitivity and specificity you are quoting was derived from only patients with cancer who had red, swollen, painful legs or whether the study in- cluded patients who had no leg symptoms and who were being evaluated for other conditions (eg, possible pul- monary embolism). If you are careful to use only sensi-

clinically similar to your own pati the results of your clinical practice mathematical models predict. An applying mathematical models in

SUGGESTED READINGS Gallagher EJ. Evidence-based evalu Ann Emerg Med 1996;28:347Ð9. Lachs MS, Nachamkin I, Edestein PH the evaluation of diagnostic tests: less stick test for urinary tract infection.A 135Ð40. Ransohoff DF, Feinstein AR. Proble in evaluating the efficacy of diagno 1978;299:926Ð30. Wheeler HB, Hirsh J, Wells P, And tests for deep vein thrombosis. Clin

Copyright 2001 by Turner White Communications Inc., Wayne, PA.

24 Hospital PhysicianNovember 2001

All rights reserved.

www.turnerwhite.com