While important progress has been achieved in our understanding of the clinical and neuroimaging characteristics, as well as genomic and neurobiological substrates of Alzheimer?s disease (AD) and related dementias, identifying individuals at a preclinical stage remains a vital priority and substantial challenge. The early and accurate identification of at-risk individuals becomes particularly important as we embark on next- generation randomized clinical trials to prevent or delay the onset of AD. The NIA recently convened a workshop involving experts from academia, nonprofit organizations, and industry with the goal to consider cost- effective strategies to improve the early detection of cognitive decline. A key recommendation by workshop participants emphasized opportunities to leverage existing longitudinal studies and apply machine learning techniques as a cost-effective approach to detect early cognitive decline. The current proposal represents a targeted step toward achieving those recommendations by leveraging a large and ongoing longitudinal study of APOE genotype and cognition (led by Dr. Richard Caselli at Mayo Clinic Arizona), and applies machine learning to identify individuals at risk for incident MCI at the earliest possible detectable stage. Machine-learning (ML) techniques implement predictive algorithms to find optimal mathematical and computational solutions to a set of complex problems. In dementia research, ML and pattern detection algorithms have been applied mainly to neuroimaging data, or neuroimaging data combined with clinical and genetic data, to distinguish prevalent MCI or AD cases from healthy controls. However, it remains to be determined whether ML algorithms can be marshalled as a key strategic and predictive approach to identify cognitively normal individuals at the earliest detectable stage of incipient decline. To address this gap in knowledge, the proposal aims to: (1) Investigate whether subtle variations among standard clinical and cognitive measures at baseline are associated with subsequent decline and incident MCI. Methods to achieve this aim consist of an ensemble ML approach, anchored by a random forests learning algorithm, applied to baseline demographic, clinical, and cognitive data as well as APOE genotype in a cohort of 784 adults. (2) Develop a probabilistic algorithm that predicts out-of-sample incident MCI cases.
This aim will be accomplished by selecting 80% of the longitudinal data as an in-sample subset and using a dynamic Bayesian network approach to model the probabilistic trajectories of diagnosis at each study visit. With this Bayesian model, it will be possible to develop an algorithm to estimate each person?s unique risk of future MCI diagnosis. (3) Validate the predictive diagnostic algorithm using the remaining 20% of longitudinal data (out-of- sample subset), plus data from additional accruals during the intervening period. Successful completion of this project will be relevant to public health by helping to improve the prediction of dementia at a very early stage using standard clinical measures, and potentially aid in the implementation of therapeutic interventions.
As our scientific field embarks on next-generation preclinical trials to prevent or delay the onset of Alzheimer?s disease, it become critically important to identify accurately who amongst cognitively healthy adults is at an increased risk of cognitive decline and development of dementia. Consistent with the recommendations from the NIA to find cost-effective solutions that capitalize on existing data and apply new methods from artificial intelligence, the current proposal aims to apply machine learning techniques to identify, at the earliest possible detectable stage, those individuals who are at increased risk of Alzheimer?s disease. Successful completion of this project will be relevant to public health as it may help to improve the prediction of dementia at a very early stage using standard clinical measures, and potentially aid in the implementation of therapeutic interventions.
