(February 18, 2004) As epidemiologists, Dr. Graham Colditz and his colleagues at the Harvard School of Public Health can make pretty accurate predictions about health outcomes in large groups of people. Can they, though, predict what will happen to specific individuals? The answer, according to Dr. Colditz, is a qualified yes: risk prediction models derived from prospective cohort studies perform fairly well, but there is room for improvement.

The basics of risk prediction
Risk prediction models are statistical models designed to identify individuals at increased risk of a given outcome. Although they have historically been used to predict outcomes in patients who already have disease, such models are now being used to predict whether a given individual will develop disease. The performance of these models is evaluated on two criteria: 1) how well the model can predict the incidence of disease in a given population; and 2) how well the model can discriminate between individuals with different outcomes.

Predicting cancer risk
Several models have been developed to predict breast cancer risk and thereby identify women who might benefit from chemoprevention, early screening, or genetic testing. Perhaps the most familiar is the Gail model, which is based on six established risk factors: age, age at first menstrual period, age at first birth, number of previous breast biopsies, history of atypical hyperplasia, and number of affected first-degree relatives. This model has been widely used to identify potential candidates for tamoxifen, yet an evaluation within the Nurses' Health Study suggests that it predicts individual risk only slightly better than a coin flip would. The odds of correctly predicting whether a given woman will develop breast cancer in the next five years are 50 percent with a coin flip and 58 percent with the Gail model.

Drs. Colditz and Rosner have developed a breast cancer model that uses additional risk factors, such as alcohol intake, weight, height, age at menopause, and use of postmenopausal hormones. When evaluated in the Nurses' Health Study, the odds of correct prediction with this model were slightly better than the Gail model: 64 percent. Dr. Colditz and colleagues have also developed ovarian cancer and melanoma models, which yield correct predictions 60 to 67 percent of the time.

Using risk prediction models to educate the public
In a unique effort to bring risk prediction models to the public, Dr. Colditz and others at the Harvard Center for Cancer Prevention have developed an interactive website called Your Cancer Risk (http://www.yourcancerrisk.harvard.edu) that incorporates risk prediction models for 12 types of cancer. These models differ significantly from previous models because the risk assessments use only dichotomized responses and are therefore much less detailed. For example, in the breast cancer assessment, individuals indicate their age at first birth as either before or after age 35. In contrast, other models for breast cancer examine this variable continuously, assigning a slight increase in risk for every year that goes by before a woman gives birth.

To examine the accuracy of the simplified assessments in Your Cancer Risk, Dr. Daniel Kim evaluated the predictive ability of three models used in the website: colon, ovarian, and pancreatic cancer. These models were able to predict the 10-year incidence of disease fairly well, and the odds of correct prediction ranged from 59 percent to 72 percent.

Conclusion
Overall, Dr. Colditz’s experiences suggest that risk prediction models for cancer can be developed efficiently. Although these models are difficult to validate, given the low 5- and 10-year risks of cancer, evaluation within large cohort studies suggests that these models can predict individual risk with fair accuracy. What remains under investigation is the clinical usefulness of individual risk prediction.

written by Catherine Tomeo Ryan