The Rationale for Evidence-Based Medicine

Diagnosis can be defined as the process of using the history, physical examination, laboratory, imaging studies, and other tests to identify the disease responsible for the patient's complaint. The advantage of knowing the responsible disease, and assigning a label to it, is that we can now make an informed decision about treatment and give our patients more accurate information about prognosis.

There are several types of reasoning used by physicians during the diagnostic process. They are:

  • algorithmic -- using flowcharts and algorithms
  • pattern-recognition -- "instant recognition" of a disease
  • exhaustive -- gathering every possible piece of data
  • hypotheticodeductive -- generating and rejecting hypotheses as more data are collected

Each is appropriate for certain situations, and inappropriate for others. For example, an algorithmic approach is useful for conditions where the information we collect from patients is discrete and accurate, such as diagnosis of anemia or an abnormal liver function test. Pattern-recognition is commonly used by primary care physicians for the diagnosis of common conditions such as urinary tract infection or sinusitis, and is a very efficient style of diagnosis. Finally, the exhaustive approach is appropriate for unusual presentations of illness, where the other modes of decision-making have failed. It is commonly used at tertiary and quaternary care centers, where patients have already received a basic evaluation.  However, it is generally inappropriate and inefficient in the primary care setting for an initial evaluation.

The hypotheticodeductive style involves proposing a differential diagnosis, asking a question, using the answer to refine the differential diagnosis, asking another question, again refining the differential, and so on until a final working diagnosis is obtained. It is generally held up as the ideal, although each of the first three approaches are perfectly appropriate under certain conditions. 

For each of these styles of diagnosis, it is important to accurately use the information which we gather from patients. That means understanding how much each symptom, sign, or test result increases or decreases the likelihood of a given disease. This process is called "revising the probability of disease", and unfortunately, it's something that physicians aren't very good at! Consider the following example:

Your 45 year old patient has a mammogram. The study is interpreted as "suspicious for malignancy" by your radiologist. Your patient asks you: "Does this mean I have cancer?", and you (correctly) answer "No, we have to do further testing." Your patient then asks, "OK, I understand that the mammogram isn't the final answer, but given what we know now, what are the chances that I have breast cancer?".  Assume that the overall risk of breast cancer in any 45 year old woman, regardless of mammogram result, is 1%. Assume also that mammography is 90% sensitive and 95% specific. Then, select your answer below:

                    1%          15%          60%          85%          95%

Were you surprised by the answer? Don't feel bad: all but one of a group of German internists (as compulsive a group as exists in nature, I dare say), given a similar question were wrong!

So, how can we make our use of diagnostic test results more rational and accurate? The answer lies in understanding concepts like sensitivity, specificity, predictive value, and likelihood ratios.