Cancer risk model requirements for counseling, prevention and early
detection
Mitchell H. Gail
Biostatistics Branch, Division of Cancer Epidemiology and Genetics
National Cancer Institute, NIH, HHS
Bethesda, MD 208927244
“Cancer risk model” refers to a model that predicts the
absolute risk of developing cancer over a defined age interval, which is
reduced by competing causes of mortality. The usefulness of a risk model
depends not only on the properties of the model, but also on the available
interventions and the intended applications. For most applications the model
should be “well calibrated”, namely should correctly predict the probability
that a person with given characteristics develops cancer in a given age
interval. In addition to good calibration, model performance is often
characterized by its “discriminatory accuracy”, which is the probability
that a randomly selected person who develops cancer has a higher risk than a
randomly selected person who does not. Although this general criterion has
some value, it is preferable to consider the benefits and losses associated
with specific applications of the risk model to assess its usefulness. Some
applications require well calibrated models but do not require high
discriminatory accuracy. Examples include: providing general perspective on
risk in counseling; deciding whether a woman in her forties might benefit
from mammographic screening; assessing the burden of disease from specific
risk factors in a population and gauging the potential population benefit of
reducing exposure to that risk factor; and sample size calculations for
designing prevention trials.
Some applications require high discriminatory accuracy. An ideal preventive
intervention would be so safe that it could be applied broadly without the
need for risk models. However, if the preventive intervention has serious
adverse effects, it should be given only to the subset of the population at
high enough risk that the expected benefit from prevention of the main
outcome outweighs the expected losses from adverse events. Such applications
require risk models with high discriminatory accuracy for high public health
impact if the targeted subset is small. Models with intermediate
discriminatory accuracy can contribute to other applications, such as
riskbased allocation of preventive resources, improved screening
recommendations compared to agebased screening, and weighing risks and
benefits of preventive interventions for individuals.
I shall discuss the usefulness of available and foreseeable models for
breast cancer in such applications, taking into account likely improvements
from including mammographic density and polygenic risk scores in breast
cancer risk models.
