Researchers in the United States have developed a new model to predict women’s risk of developing breast, uterine and ovarian cancer, based on individual lifestyle factors. These three cancers make up one-third of all invasive cancers diagnosed among Australian women, with more than 17,000 diagnosed each year.
While the prognosis for women who are diagnosed with these cancers is improving, due in part to earlier detection and more effective treatment, nearly 4,000 Australian women died due to one of these cancers in 2007. Added to this is the impact a cancer diagnosis has on quality of life and psychological well-being.
The ideal way of reducing the impact of cancer is to prevent the development of cancer in the first place. An estimated one-third of cancers can be prevented through improvements in lifestyle behaviours, such as not smoking, reducing alcohol intake, increasing physical activity and having a more nutritious diet.
So, how does the new tool work, and how accurate is it likely to be?
Assessing risk
The researchers used the results of a large study of almost 200,000 healthy women aged over 50 to develop mathematical models that tries to estimate a woman’s risk of each type of cancer. They then analysed a separate cohort of 64,440 initially healthy women aged 55 to see how the model performed.
The model is based on the combination of known risk factors a woman has, including body mass index, smoking status and level of alcohol consumption. It also takes into account other factors such as:
- age at birth of first child
- number of children
- family history of breast of ovarian cancer
- age at menopause
- the use of oral contraceptives and hormone replacement therapy.
The magnitude and direction of these associations depend on the specific cancer.
The idea behind these models was that having an estimate of future risk would help women and their doctors make more informed decisions about cancer screening, prophylactic surgery, improving preventive behaviours, and use of specific medicines.
Although there have been previous risk factor models for breast cancer, this is the first study that has examined models for breast cancer, endometrial cancer and ovarian cancer from the one cohort.
There are two ways in which this tool could be used. First, the relative risk estimates show the impact each risk factor has on subsequent risk, and so could provide motivation of modify those risks. A woman who drinks at least one alcoholic drink per day has about a 25% greater risk of endometrial cancer than a teetotaller. So in terms of endometrial cancer risk, a woman is better off not drinking alcohol.
The second use is to estimate, based on a specific combination of risk factors, a woman’s risk (or probability) is of being diagnosed with one of these cancers in the next 20 years. These absolute estimates are often provided through web-based dissemination tools, such as this one for cardiovascular disease.
At present, it seems the researchers of the tool have only made computer codes available for statistical software.
The down sides
The model has some shortcomings. These studies only include Caucasian women aged 50 years and over. So we cannot extrapolate the results to younger women, nor to women of other races or nationalities.
Also, the cohorts did not include all eligible women. The response rate for the original study cohort, for example, was only 17.6% (approximately 570,000 out of 3.5 million people).
It is possible that the people who responded to the study had different characteristics to those who didn’t. They might have been healthier or more interested in health issues, both of which may impact on their risk of cancer and so on the final results.
The authors also acknowledge the models aren’t intended to predict the probability of the three cancers among women at much higher risk, such as those with the BRCA1 or BRCA2 mutations.
How do the models perform?
Performance is typically measured in two ways – calibration and discrimination.
Calibration measures how well the model, developed in the initial cohort, corresponds to the real outcome possibility: in this case, the observed proportion of cancer diagnoses in the second cohort.
In contrast, discrimination measures the ability of the model to predict a given outcome (that is, a cancer diagnosis) based on the combination of risk factors for a specific person.
In terms of calibration, the predictive risk model performed well. Although it over-estimated the overall risk of endometrial cancer by about 20%, the predicted risk for breast and ovarian cancers agreed closely with the observed risk.
Discrimination, however, was not so impressive. When considering that a value of 0.5 is analogous to tossing a coin, and 1.0 is perfect prediction, the discrimination values for breast cancer and ovarian cancer were 0.58 and 0.59 respectively. The predictive capacity for endometrial cancer was slightly better at 0.68.
What these results mean in practice?
These models provide important insights into the absolute burden of these cancers in a population of white women aged 50 years and over. It’s also useful in gauging the potential for how the population risk of cancer could be reduced using preventive interventions among this population.
But it’s not very effective at predicting individual woman’s risk of cancer. While the factors included in the statistical models explain some of the risk, there remain other unmeasured factors that dictate whether a woman will get one of these cancers.
This means that a woman might have none of the risk factors and still be diagnosed with breast cancer, or she might have all of the risk factors and not be diagnosed in the next 20 years. Given that there is still a lot to be learnt about the causes of most types of cancer, this conclusion is not really surprising.
The take-home message of this research is that there are measures by which women can reduce their risk of getting these types of cancer, and hopefully these results increase their motivation to do so.