Predicting Alzheimer's risk: why and how?

Because the pathologic processes that underlie Alzheimer's disease (AD) appear to start 10 to 20 years before symptoms develop, there is currently intense interest in developing techniques to accurately predict which individuals are most likely to become symptomatic. Several AD risk prediction strategies - including identification of biomarkers and neuroimaging techniques and development of risk indices that combine traditional and non-traditional risk factors - are being explored. Most AD risk prediction strategies developed to date have had moderate prognostic accuracy but are limited by two key issues. First, they do not explicitly model mortality along with AD risk and, therefore, do not differentiate individuals who are likely to develop symptomatic AD prior to death from those who are likely to die of other causes. This is critically important so that any preventive treatments can be targeted to maximize the potential benefit and minimize the potential harm. Second, AD risk prediction strategies developed to date have not explored the full range of predictive variables (biomarkers, imaging, and traditional and non-traditional risk factors) over the full preclinical period (10 to 20 years). Sophisticated modeling techniques such as hidden Markov models may enable the development of a more comprehensive AD risk prediction algorithm by combining data from multiple cohorts. As the field moves forward, it will be critically important to develop techniques that simultaneously model the risk of mortality as well as the risk of AD over the full preclinical spectrum and to consider the potential harm as well as the benefit of identifying and treating high-risk older patients.


Abstract
Because the pathologic processes that underlie Alzheimer's disease (AD) appear to start 10 to 20 years before symptoms develop, there is currently intense interest in developing techniques to accurately predict which individuals are most likely to become symptomatic. Several AD risk prediction strategies -including identifi cation of biomarkers and neuroimaging techniques and development of risk indices that combine traditional and nontraditional risk factors -are being explored. Most AD risk prediction strategies developed to date have had moderate prognostic accuracy but are limited by two key issues. First, they do not explicitly model mortality along with AD risk and, therefore, do not diff erentiate individuals who are likely to develop symptomatic AD prior to death from those who are likely to die of other causes. This is critically important so that any preventive treatments can be targeted to maximize the potential benefi t and minimize the potential harm. Second, AD risk prediction strategies developed to date have not explored the full range of predictive variables (biomarkers, imaging, and traditional and non-traditional risk factors) over the full preclinical period (10 to 20 years). Sophisticated modeling techniques such as hidden Markov models may enable the development of a more comprehensive AD risk prediction algorithm by combining data from multiple cohorts. As the fi eld moves forward, it will be critically important to develop techniques that simultaneously model the risk of mortality as well as the risk of AD over the full preclinical spectrum and to consider the potential harm as well as the benefi t of identifying and treating high-risk older patients.
Another recent study extended this work by noting that a tremendous range of factors appear to predict AD, suggesting that AD is more likely in people with a broad decline in health. Song and colleagues [7] found that 19 non-traditional risk factors such as a general health question ('How good is your health?') and sensory questions ('How good is your eyesight or hearing?') predicted which individuals were likely to develop AD, even after accounting for traditional risk factors such as age, sex, education, cognitive function, and cardiovascular disease. Implicit in this study is the assumption that, before leading to symptoms of cognitive impairment, AD is associated with non-specifi c symptoms such as worse self-rated health. However, the prognostic accuracy of this model was only fair (c statistic = 0.66). In fact, in another study, demographics alone were found to be more predictive of dementia risk (c statistic = 0.72), and signifi cant increases in prognostic accuracy were associated with the addition of APOE genotype (c statistic = 0.75) and vascular risk factors (c statistic = 0.79) [8].
Taken together, these studies suggest that there are a variety of approaches for identifying with fair to moderate levels of discrimination those individuals who are at high risk of developing AD. However, all of these models suff er from two important limitations in their approach to the prediction of AD risk.
First, none of these models explicitly accounts for the competing risk of death. AD is predominantly a disease of older adults who experience high rates of death from other causes. To eff ectively target preventive inter ventions -in particular, pharmacologic interventions that may have adverse side eff ects -AD prediction models must identify individuals who are likely to suff er symptomatic AD before death. Many risk factors for AD are also risk factors for mortality (for example, older age, vascular risk factors, and functional limitations) [9]. Th erefore, some patients who are at high risk for AD may be at even higher risk for death before symptomatic AD. To appropriately balance the potential benefi t of preventive intervention with the potential harm, AD prediction models must account for death and identify individuals whose risk of AD outweighs their risk of death.
Second, dementia risk models developed to date have been developed in cohort studies that focused on a limited range of potential predictors and had relatively short follow-up periods (<10 years) or relatively narrow risk windows (for example, mid-life only and late-life only). Th us, these studies have not examined the full range of AD predictors (biomarkers, neuroimaging techniques, and traditional and non-traditional risk factors) over the full 10-to 20-year preclinical risk period. Th e ideal study would include repeated ascertainment of thousands of individuals over several decades, requiring a tremendous investment in resources. Furthermore, these prospectively collected data would not be available for decades, suggesting that alternative modeling techniques need to be employed. One possible solution is to use hidden Markov models [10], which potentially could be used to combine data from multiple sources to model disease state transitions across the full AD clinicalpathologic spectrum.
In conclusion, prediction of AD risk is a relatively new fi eld of inquiry. Several alternative approaches with moderate levels of accuracy have been developed, but none is ready for widespread clinical use. As the fi eld moves forward, it will be critically important to develop techniques that simultaneously model the risk of mortality as well as the risk of AD over the full preclinical spectrum and to consider the potential harm as well as the benefi t of identifying and treating high-risk older patients.