Risk Prediction Versus Diagnosis: Preserving Clinical Nuance in a Binary World

IN RESPONSE:

Dr. Warner, Dr. Djulbegovic, and Dr. Patrick each point to practical problems with a prediction approach. Dr. Warner concurs with a point we made in our article, which is that use of discrete categories (disease vs. no disease) simplifies clinical management and communication. We also agree with him that changes in medical education are needed to help physicians understand and communicate the results of risk prediction. A point of disagreement is that physicians' and patients' poor understanding of probabilities is a problem specific to the risk-prediction approach. For example, even if we use binary diagnostic categories, we would still want to inform the patient about their risk (“Mr. Jones, you have hypertension, which means a 20% risk for having a heart attack”). Conversely, we might use prediction models without reference to numbers at all (“Mr. Jones, you are at high risk for a heart attack, so I am going to write you a prescription for some pills”).

Dr. Djulbegovic argues that regardless of whether we use a binary diagnostic category or a risk prediction model, we still have to choose a threshold to treat a patient. This can cause problems when results are close to the threshold. We agree that there is room for both descriptive and normative research on decision making near decision thresholds. We also agree with Dr. Patrick's point that we currently live in a binary world and enjoyed his description of the numerous ways in which those outside the examination room force a physician to think in simple binary terms. We are not naive about the practical challenges of implementing a prediction approach. That said, we must make medical progress in the best interests of our patients and hope that outside forces and structures follow along: We would certainly hate to see, for example, the military's need for specific criteria for service disqualification affect the way we practice medicine.

Dr. Swerlick makes a distinction between having symptoms or functional impairment and having only a risk factor for a disease. Although we focused on risk factors, we believe that binary diagnostic thinking is often inappropriate for symptomatic disease. For example, many people have symptoms of depression; a choice of a particular cut-point on a spectrum of severity does not create 2 natural categories of depressed and not depressed. A prediction approach would focus on whether treatment would do more good than harm.