Nomograms

Deriving Outcome Probabilities

All medical decisions are based on the predicted probability of different outcomes. Imagine a 35-year-old patient who presents to a physician with a 6-month history of cough. A doctor in Chicago might recommend a test for asthma, which is a common cause of chronic cough. If the same patient presented to a clinic in rural Africa, the physician might be likely to test for tuberculosis. Both physicians might be making sound recommendations based on the predicted probability of disease in their locale. These physicians are making clinical decisions based on the overall probability of disease in the population. These types of decisions are better than arbitrary treatment but treat all patients the same.

A more sophisticated method for medical decision making is risk stratification. Physicians will frequently assign patients to different risk groups when making treatment decisions. Risk group assignment will generally provide better predicted probabilities than estimating risk according to the overall population. In the previous cough example, a variety of other factors might impact the predicted risk of tuberculosis (e.g., fever, exposure to tuberculosis, and history of tuberculosis vaccine) that physicians are trained to explore. Most risk stratification performed in clinical practice is based on rough estimates that simply order patients into levels of risk, such as high risk, medium risk, or low risk. Nomograms provide precise probability estimates that generally make more accurate assessments of risk.

A problem with risk stratification arises when continuous variables are turned into categorical variables. Physicians frequently commit dichotomized cutoffs of continuous laboratory values to memory to guide clinical decision making. For example, blood pressure cut-offs are used to guide treatment decisions for hypertension. Imagine a new blood test called serum marker A. Research shows that tuberculosis patients with serum marker A levels greater than 50 are at an increased [Page 905]risk for dying from tuberculosis. In reality, patients with a value of 51 might have similar risks compared with patients with a value of 49. In contrast, a patient with a value of 49 would be considered to have the same low risk of a patient whose serum level of marker A is 1. Nomograms allow for predictor variables to be maintained as continuous values while allowing numerous risk factors to be considered simultaneously. In addition, more complex models can be constructed that account for interactions.

...