Machine learning amplifies the effect of parental family history of Alzheimer's disease on list learning strategy.

Identification of preclinical Alzheimer's disease (AD) is an essential first step in developing interventions to prevent or delay disease onset. In this study, we examine the hypothesis that deeper analyses of traditional cognitive tests may be useful in identifying subtle but potentially important learning and memory differences in asymptomatic populations that differ in risk for developing Alzheimer's disease. Subjects included 879 asymptomatic higher-risk persons (middle-aged children of parents with AD) and 355 asymptotic lower-risk persons (middle-aged children of parents without AD). All were administered the Rey Auditory Verbal Learning Test at baseline. Using machine learning approaches, we constructed a new measure that exploited finer differences in memory strategy than previous work focused on serial position and subjective organization. The new measure, based on stochastic gradient descent, provides a greater degree of statistical separation (p = 1.44 × 10-5) than previously observed for asymptomatic family history and non-family history groups, while controlling for apolipoprotein epsilon 4, age, gender, and education level. The results of our machine learning approach support analyzing memory strategy in detail to probe potential disease onset. Such distinct differences may be exploited in asymptomatic middle-aged persons as a potential risk factor for AD.