Hierarchical or mixed-effects models have been used to model longitudinally collected data from individuals or clusters from many diverse areas of agricultural, biological and medical sciences. Recently, mixed models along with Bayes’ theorem have provided a method for predicting future events based on the corresponding distribution of longitudinal trends from variables measured over time, and the calculated posterior probabilities of the occurrence of a given event. The method is illustrated by examining the ability of numerous behavioral, physiological, and biochemical measures collected over the entire adult lifespan of individuals participating in the Baltimore Longitudinal Study of Aging (BLSA) to predict diagnosis of incident AD. From 1576 participants aged 20-97 years and initially free of AD from the BLSA, we applied our longitudinal prediction method to generate posterior probabilities for AD for each of 15 candidate measures separately by sex and APOE ε4 carrier status.
During mean follow-up of 14.4 years, 6.7% of participants were diagnosed with AD. All candidate measures predicted AD, but with varying degrees of accuracy (represented by area under the receiver operating characteristic curve, AUC) and mean lead times (MLT). For women ε4 non-carriers, LDL was the most accurate predictor (AUC = 0.80, MLT = 10.7 years), followed by total cholesterol, whereas for women ε4 carriers the most accurate predictors were depressive symptoms (AUC = 0.78, MLT = 9.1 years), followed by systolic blood pressure (SBP). For men ε4 non-carriers and carriers, mean arterial pressure was the most accurate predictor (AUC = 0.71, MLT = 8.4 years; and AUC = 0.76, MLT = 8.2 years). In general, lipids were better predictors in APOE ε4 non-carriers and BP measures in APOE ε4 carriers. Age modified the effects of depressive symptoms, SBP, and LDL on AD probability. In conclusion, longitudinal changes in blood pressure, lipid levels, and depressive symptoms are variably powerful predictors of AD diagnosis, depending on age, sex, APOE status. Widely available clinical measures collected longitudinally may be useful for early prediction of AD.