Inferring Alzheimer’s disease pathologic traits from clinical measures in living adults

Background and Objectives: Develop imputation models using clinical measures to infer Alzheimer's disease neuropathologic changes (ADNC) in living adults to identify adults at risk for Alzheimer's disease (AD). Methods: We used clinical and postmortem data from two prospective cohort studies, Memory and Aging Project (MAP) and Religious Orders Study (ROS). We used generalized linear regression models with Elastic Net penalty to train imputation models of ADNC traits (beta Amyloid, tau tangles, global AD pathology, and NIA Reagan), in MAP decedents using clinical measures collected at the last visit as predictors. ROS cohort was used as an independent validation and test data. We validated these models in ROS decedents and applied the models to baseline clinical data of ROS participants to infer baseline ADNC traits. Baseline clinical data were collected an average of 8 years before the last follow up. We used Cox proportional hazard models to test if inferred baseline ADNC traits predicted incident AD dementia (ADD). In addition, two sample t tests were used to examine if inferred baseline ADNC traits predicted adults with a high risk of pathologic AD profiled at death. Results: By applying imputation models to clinical measures collected at the last visit in ROS to validate the imputation models, we obtained predicted R2 as 0.188 for beta Amyloid, 0.316 for tau tangles, and 0.262 for global AD pathology. The prediction area under the receiver operating characteristic curve (AUC) for the dichotomous NIA Reagan was 0.765. All four inferred ADNC traits at the last visit strongly discriminated postmortem NIA Reagan status (pvalues <10E-28). The inferred baseline levels of all four ADNC traits predicted ADD, with higher accuracies for predicting ADD in Year 3 (AUC ranging in 0.861 and 0.919) versus Year 5 (AUC from 0.842 to 0.896), and the highest accuracy was obtained using inferred NIA Reagan status. The inferred baseline levels of all four ADNC traits significantly discriminate individuals with postmortem pathologic AD (all pvalues < 1.5E-7). Conclusions: Inferred baseline levels of ADNC traits derived from clinical measures discriminate adults at risk for ADD and pathologic AD profiled at death. Further studies are needed to determine if repeated measures of inferred ADNC traits can be used to monitor the accumulation of ADNC traits during the prolonged course of AD.